Learning Library

← Back to Library

Weekend AI News: Encryption, Orion, Gemini

Key Points

  • Apple is exploring homomorphic encryption so that images can be processed on its servers without ever being decrypted on the device, allowing secure, privacy‑preserving visual recognition.
  • A weekend rumor dubbed “Orion” claimed OpenAI’s next model would be 100× more powerful and launch in November, but OpenAI publicly denied any such release schedule.
  • Despite OpenAI’s denial, industry observers expect the company to debut a new model by December, given the pattern of end‑of‑year releases.
  • Google is reportedly gearing up for a major Gemini update slated for December, signaling a competitive push in the AI model space.

Full Transcript

# Weekend AI News: Encryption, Orion, Gemini **Source:** [https://www.youtube.com/watch?v=hZ-9ks8R15Y](https://www.youtube.com/watch?v=hZ-9ks8R15Y) **Duration:** 00:07:15 ## Summary - Apple is exploring homomorphic encryption so that images can be processed on its servers without ever being decrypted on the device, allowing secure, privacy‑preserving visual recognition. - A weekend rumor dubbed “Orion” claimed OpenAI’s next model would be 100× more powerful and launch in November, but OpenAI publicly denied any such release schedule. - Despite OpenAI’s denial, industry observers expect the company to debut a new model by December, given the pattern of end‑of‑year releases. - Google is reportedly gearing up for a major Gemini update slated for December, signaling a competitive push in the AI model space. ## Sections - [00:00:00](https://www.youtube.com/watch?v=hZ-9ks8R15Y&t=0s) **Apple’s Homomorphic Encryption Demo** - The speaker explains how Apple could transmit an image encrypted with homomorphic encryption to its servers, allowing the cloud to run machine‑learning calculations (e.g., identifying the Eiffel Tower) on the ciphertext without ever exposing the raw image. ## Full Transcript
0:00five quick pieces of AI news from the 0:02weekend I hope you have your coffee 0:04because the first one uses the phrase 0:05homomorphic encryption and if you're 0:08wondering what is that the basic idea is 0:11you can put something into what's called 0:12CER text or encrypted strings and it 0:15retains computationally the same 0:18structure as the original uh piece of 0:20data which means you can do math on it 0:23which is important because machine 0:24learning is basically math and so the 0:27example that Apple gave was basically 0:29what if you take a picture of something 0:31and you want to find out what that 0:32something is which is a very classic 0:34thing that you ask uh open AI I I 0:37actually asked open AI what was in an 0:38image just this weekend um well let's 0:41say you ask Apple intelligence that 0:42instead let's say you're on your phone 0:44and you ask Apple intelligence what's in 0:46this 0:47image and you send the image or apple 0:51sends the image encrypted in homomorphic 0:55encryption to its servers to find out so 0:58the actual image is UN encrypted doesn't 1:00leave your phone you're you're the image 1:03is secure on your phone and the 1:07homomorphic encryption is the only thing 1:09that 1:10departs at that 1:12point Apple then does computations on it 1:15to find other homomorphically similar 1:18images what's weird is I can say things 1:21like they do math on it because it's the 1:23same shape as the data but what that 1:25means in practice is it can tell if 1:27you're looking at the Eiffel Tower so if 1:29you're looking at the Eiffel Tower and 1:30you say what is this image apple can 1:32come back because of homomorphic 1:34encryption and say that's the Eiffel 1:37Tower based on other computationally 1:39similar images in the database all of 1:41which remain 1:43encrypted it's weird it works I can't 1:47quite explain how it works either it it 1:49seems odd to me that it should work that 1:51way but homomorphic encryption is very 1:53cool we'll have to do a separate video 1:54on it okay piece of news number two a 1:58ton of discussion over the weekend on 2:00Orion Orion is the code name or the 2:03rumored code name for open ai's new 2:07model The Verge did a major report 2:10claiming that it was going to be a 100 2:11times more powerful and it was going to 2:13debut in November which matches the 2:16rumored timeline of a lot of other 2:17Advanced models this end of year and 2:20they were saying specifically it's so 2:22powerful it's going to companies first 2:25we're not sure when it's going to be 2:26released to 2:28individuals and then immediately open AI 2:31push back still over the weekend and 2:33said no no no no no no we are not 2:35releasing a model called Orion in 2:40November now that leaves them a lot of 2:42wiggle room they could be releasing a 2:44different model they could be releasing 2:46a differently named model or they could 2:49be choosing to release it in December 2:51nobody's clarifying we don't know so 2:55stay tuned but I would still expect open 2:57AI to release something in December if 2:59only because you were seeing other 3:02releases come along and that brings me 3:04to piece of news number three which is 3:07that Google is preparing a major Gemini 3:11release in December rumor also broke 3:13over the weekend and will 3:16include the proposal or the uh enabling 3:21apis or whatever you want to put it it's 3:23very unclear it's very it's another sort 3:25of on the edge of news and rumor uh they 3:28will include an enablement for a Google 3:31AI agent driven by 3:33Gemini that's the 3:35tldr I noticed that this particular 3:38piece of news dropped after the open AI 3:40Verge article and I do wonder how much 3:44we are at this point in a news and rumor 3:46cycle where everyone wants to tell you 3:48that their that their particular model 3:51is going to be ready right at the right 3:52time and it's going to be super 3:55powerful I still think we're going to 3:57get a big model release before the end 3:58of the year I I think it will probably 4:00be somewhere in the December 1 to 15 4:02time frame we will 4:05see okay uh the next piece of news is 4:09coinbase coinbase is also getting into 4:11the mix but they of course don't have a 4:13major llm uh model that they're going to 4:16be building on instead what they're 4:18going to be doing is just putting AI 4:22agents autonomously on the blockchain 4:24and that's been their goal uh or so they 4:26say and so they are putting out a press 4:29relas saying that's what they're doing 4:30we'll see when something actually comes 4:32out 4:34and when when they release 4:38that I just keep noticing they are 4:40releasing that after the AI meme Coin 4:44News broke over the last three or four 4:46weeks you remember I talked about the AI 4:48meme coins I talked about gosias Maximus 4:50there's been other AI mem coins this is 4:53the moment when coinbase is like ah yes 4:55we're going to do autonomous AI agents 4:57on blockchain 5:01we will see sounds like it needs some 5:02governance to 5:04me 5:06okay almost last second to last but not 5:09least uh so piece of news number five uh 5:13Nvidia is getting exported to Russia via 5:18an Indian pharmaceutical com company 5:20that seems to be a coverup called 5:23Shrea and the reason this is news is 5:26because the h100 chips which are not 5:28actually The Cutting Edge chips now 5:29they're like one generation back are not 5:32allowed to be exported to certain 5:35countries and that includes Russia and 5:38Russia is figuring out how to import 5:39them anyway apparently at scale like 5:42they've imported 1,00 or so Dell servers 5:46with h100 chips uh under cover of this 5:49pharmaceutical company in India and 5:51they're doing it uh presumably to set up 5:53their own uh AI modeling data centers we 5:57will see 6:00finally uh there was a major report on 6:03the gender gap in AI on average it found 6:06over uh multiple studies uh and I looked 6:10at the chart like it really was like 20 6:12different studies on AI usage 25% less 6:16usage of AI by women versus 6:20men and this was not just us Tech 6:22workers in fact US tech workers if 6:24you're wondering were was the only 6:27surveyed demographic 6:29where women used AI more than men uh 6:32although it was virtually identical so 6:34that's sort of a nominal difference um 6:36but US tech workers use AI apparently 6:38two-thirds of us use AI all the time 6:41um but the gender gap is something that 6:45did persist uh across just about every 6:48other demographic that that I was 6:49looking at and I looked into the paper a 6:51little bit super interesting no one can 6:54really explain why uh and it's something 6:58that since AI usage is correlated with 7:01AI fluency it's concerning it's it's 7:03well it's concerning regardless so there 7:06you go that's the last piece of AI news 7:08I hope you enjoyed it I hope you had 7:10coffee I hope you can pronounce 7:12homomorphic encryption