AI salaries are outpacing NBA MVPs. Grok is popping heads, and stirring controversy, as competitors amongst prime AI labs heats up.
This week, Mike and Paul unpack OpenAI’s replace that has the potential to show ChatGPT right into a digital assistant, Meta’s $200M+ AI expertise raids, and the spiraling drama at Grok. They break down Microsoft’s AI-driven layoffs, AI browser competitors, Apple’s quiet AI pivot, and what it actually means when superintelligence turns into low cost and all over the place.
Pay attention or watch beneath—and see beneath for present notes and the transcript.
Pay attention Now
Watch the Video
Timestamps
00:00:00 — Intro
00:05:47 — ChatGPT Agent
00:14:17 — Grok 4
00:17:46 — Grok Controversy
00:32:07 — Meta Superintelligence Labs
00:35:16 — Windsurf Drama
00:37:28 — Kimi 2
00:42:35 — AI Browsers
00:47:05 — Microsoft’s Layoffs and AI
00:52:27 — Extra Meta Updates
00:55:53 — Extra OpenAI Updates
01:03:13 — Google Updates
01:07:57 — Apple Would possibly Use Anthropic or OpenAI for Siri
01:10:59 — AI Product and Funding Updates
- Grammarly Acquires Superhuman
- NotebookLM Releases Featured Notebooks
- Considering Machines Lab Updates
Abstract:
ChatGPT Agent
OpenAI has simply given ChatGPT a serious improve: it may well now take actions in your behalf. They’re calling this new functionality “ChatGPT agent.”
This agent can carry out duties end-to-end, comparable to reviewing calendars, planning meals, researching rivals, or creating displays—with out consumer micromanagement.
The system integrates OpenAI’s earlier advances: internet interplay from OpenAI Operator, deep analysis capabilities, and ChatGPT’s conversational intelligence. Initially, the function is obtainable to Professional, Plus, and Group customers, with growth to Enterprise and Training customers coming quickly.
Grok 4
Grok 4, xAI’s newest mannequin, is being hailed as probably the most clever AI to this point.
Educated utilizing reinforcement studying on the huge 200,000-GPU Colossus cluster, Grok 4 Heavy grew to become the primary mannequin to attain over 50% on Humanity’s Final Examination, a benchmark for expert-level reasoning.
Past uncooked intelligence, Grok 4 introduces native software use, enabling it to autonomously run code, browse the net, search X, or analyze visible media—choosing the precise software for the duty and delivering nuanced, real-time insights.
Grok Controversy
Grok additionally had its fair proportion of controversy these previous couple weeks.
After a system replace, Grok started to generate a wave of antisemitic content material together with frequently praising Adolf Hitler.
This got here after a system replace that explicitly instructed Grok to “not shrink back” from politically incorrect claims, which is a immediate that has since been eliminated.
Among the many extra disturbing outputs, Grok advised Hitler could be a superb chief for contemporary America, made repeated references to Jewish final names as indicators of utmost activism and referred to itself at one level as “MechaHitler.”
xAI apologized for the difficulty, saying the issue wasn’t the language mannequin itself, however an upstream code change that “by chance reactivated outdated, deprecated directions.”
In line with this clarification, Grok was in a position, unintentionally, to basically echo extremist consumer posts as an alternative of filtering or rejecting them.
That, paired with some consumer prompts, led Grok to strengthen hate speech. The offending code has since been deleted and xAI says it’s refactored the system and added new safeguards.
This week’s episode is delivered to you by MAICON, our sixth annual Advertising and marketing AI Convention, taking place in Cleveland, Oct. 14-16. The code POD100 saves $100 on all cross sorts.
For extra data on MAICON and to register for this yr’s convention, go to www.MAICON.ai.
This episode can be delivered to you by our fiftieth Intro to AI session taking place on August 14.
You may be taught what AI is, why it issues, and discover actual use instances and trusted instruments on your work. Go to www.marketingaiinstitute/com/intro-to-ai to register.
Learn the Transcription
Disclaimer: This transcription was written by AI, because of Descript, and has not been edited for content material.
[00:00:00] Paul Roetzer: AI researchers are actually getting paid extra prime AI researchers getting paid greater than the very best paid skilled athletes. Everyone could be like, oh my God, these athletes are paid a lot cash. That is a ridiculous contract. After which right here we obtained 10 of those folks. Welcome to the Synthetic Intelligence Present, the podcast that helps what you are promoting develop smarter by making AI approachable and actionable.
[00:00:22] My title is Paul Roetzer. I am the founder and CEO of Smarter X and Advertising and marketing AI Institute, and I am your host. Every week I am joined by my co-host and advertising AI Institute Chief Content material Officer Mike Kaput, as we break down all of the AI information that issues and provide you with insights and views that you need to use to advance your organization and your profession.
[00:00:43] Be a part of us as we speed up AI literacy for all.
[00:00:50] Welcome to episode 1 58 of the Synthetic Intelligence Present. I am your host, Paul Roetzer. I am with my co-host Mike Kaput. We’re again after a pair weeks of summer season break. [00:01:00] For precise summer season break. I used to be on trip for every week and, course improvement. Mike and I’ve each form of been within the lab constructing course for the AI Academy relaunch that’s developing very quickly.
[00:01:13] We’ll in all probability have an announcement subsequent week. if you’re an AI Academy member, I might say keep tuned, for some stuff this week. You may in all probability be getting some data very, quickly about this. So I’ve, I do not know, it was humorous. There is a couple subjects we’re gonna undergo that truly in the course of constructing my programs modified the character of the programs.
[00:01:37] So, so when this stuff we launched, so I used to be particularly engaged on the ai, elementary collection, which has like, intro to ai, state of ai. AI brokers 101 AI timeline prompting 101 Gen AI 101, and it was identical to every single day one thing was taking place, it is like, wow, I gotta add that into the course.
[00:01:57] So, I do not know, I really feel like, as a result of I’ve [00:02:00] been constructing programs about these items for the final two weeks, I might, I used to be going by way of in prep this morning, Mike, by way of just like the curated listing that you just had put collectively. I am like, did not we discuss all this already? I used to be like, oh no. This was within the AI brokers 1 0 1 deck I used to be constructing is the place I talked about this.
[00:02:16] So we now have a ton to cowl after a few weeks away. I do not know, Mike, it was like 70 or 80 subjects perhaps in our, simply, simply, yeah. Weekly sandbox. So this week, greater than ever, be sure to’re subscribed to the, this week in AI publication that Mike places collectively every Tuesday, as a result of there are simply 30 or 40 articles that we might’ve cherished to have gotten to, however there’s simply no approach in a single episode to do all of it.
[00:02:44] After which I am, I am leaving once more. Tuesday morning. So we’re recording this on Monday, July twenty first. I am leaving once more tomorrow morning, so we could not do a second episode this week. I am, I am not even gonna be round to have the ability to do it. So we’re, form of sneaking this episode in, in just like the, this [00:03:00] 12 hour window we had in between course improvement and journey.
[00:03:04] okay, so that each one being stated, we’re gonna go pure fast hearth at this time. So if you’re our common listener, that we normally do three fundamental subjects the place we’ll ship, , normally seven to 10 minutes, generally quarter-hour on like a giant subject. After which we’ll do normally seven to 10 fast hearth objects.
[00:03:22] Immediately is all fast hearth. so in the event you’re new to the podcast, that is the primary time you are listening. It isn’t normally all fast hearth. However, you may, you may get a way of form of the way it all works. Positive, positive. So heaps to undergo. This episode is delivered to us by MAICON 2025. That is the Advertising and marketing Institute’s Advertising and marketing AI Convention.
[00:03:41] That is our sixth annual occasion that’s taking place October 14th to the sixteenth in Cleveland, Ohio. you possibly can go to the positioning now. It is MAICON.ai. That is MAICON.ai. You’ll be able to take a look at the agenda. It is, I do not know, like 90 ish %, that we’re nonetheless gonna be making some large bulletins [00:04:00] round normal periods and fundamental stage stuff.
[00:04:02] So keep tuned for these. However you possibly can, once more, go examine all that out. We would like to have you ever. Cleveland, which is our house base, is wonderful in October. it is my favourite time of yr truly in Cleveland within the fall. So we might like to have you ever be part of us, Mike, and I will be there. The entire crew from SmarterX and Advertising and marketing AI Institute will probably be there.
[00:04:20] ticket costs do go up on July twenty fifth. So in the event you’re listening to this, the week, this episode is dropping. get in now earlier than these costs leap. They, they normally go up, I feel, I feel it is like 100 bucks on the finish of every month or one thing, is how we do it. Alright, after which, one different fast observe.
[00:04:36] It is a free factor. We have now our intro to AI class that I began educating in October of 2021. So each month since 2021 we have been doing this free class. We have had, I feel near 40,000 folks have now registered for this class by way of, these previous couple of years. August 14th, Thursday, August 14th is our fiftieth episode.
[00:04:58] I feel the crew is [00:05:00] planning some enjoyable issues across the fiftieth episode. I do know we at the very least talked about that at one level. We’re gonna do some enjoyable stuff. So, that will be an excellent one to attend. We’ll put the hyperlink within the present notes, however you will discover that by way of Advertising and marketing Institute’s web site. and it is also in all probability linked to on Good Rx Do ai.
[00:05:17] So, be part of us for Intro to AI on August 14th for the fiftieth version of that free class. Okay? Speedy hearth away. Mike. I imply I noticed stuff already this morning. Oh my gosh. I do know. I lastly, at one level for, for Mike’s sanity, I feel I simply began placing stuff into episodes 1 59 sandbox on like Friday. I used to be like, you can not probably get the rest into 1 58.
[00:05:41] Mike Kaput: I very a lot appreciated that as a result of I used to be additionally following these being like, oh my gosh, the information would not cease.
[00:05:47] ChatGPT Agent
[00:05:47] Mike Kaput: Yeah. Alright, so first up, OpenAI has simply given ChatGPTA main improve. It will possibly now take actions in your behalf utilizing its personal digital laptop to get actual world duties finished from begin to end.
[00:06:02] They’re calling this new functionality Chat, GPT Agent and OpenAI writes in a weblog submit about this quote, ChatGPT can now do be just right for you utilizing its personal laptop dealing with advanced duties from begin to end. So they provide some examples. You may ask ChatGPT to do issues like take a look at my calendar and temporary me on upcoming consumer conferences.
[00:06:24] Plan and purchase components, as an example, to make Japanese breakfast for 4 or analyze three rivals and create a slide deck. So then Chad, gpt will intelligently navigate web sites. Filter outcomes immediate you to log in securely when wanted, run code, conduct evaluation, and even ship editable slide chosen spreadsheets that summarize its findings.
[00:06:48] Now what’s cool is it does this all by way of a unified age agentic system, which brings collectively three strengths of earlier breakthroughs on opening AI’s half. So first is operators [00:07:00] potential to work together with web sites. Second is deep analysis is ability in synthesizing data. And third is chat GT’s intelligence and conversational fluency.
[00:07:10] Now for now chat, GPT agent is obtainable to Professional Plus and crew customers and there is extra entry rolling out quickly to enterprise and training customers. So Paul, I have been enjoying round with this just a bit bit to this point. I do wanna spotlight a touch upon this that I discovered attention-grabbing. That is from Ethan Molik.
[00:07:30] He stated, quote, I had early entry and chat. GPT agent is, I feel, a giant step ahead for getting AI to do actual work. Even at this stage, it does a superb job autonomously doing analysis and assembling Excel information with formulation, PowerPoint, et cetera. It provides a way of how brokers are coming collectively. It feels far more like working with an precise human intern able to a wider vary of analytical and laptop duties.
[00:07:55] So what do you assume, Paul? Are we now lastly on the tipping level the place brokers are [00:08:00] beginning to work?
[00:08:01] Paul Roetzer: I do not know. I feel that it’s a development on the spectrum of autonomy. In order a fast reminder, so that is truly one of many ones I used to be referring to upfront. Like I used to be engaged on the AI brokers course and I used to be truly going by way of ultimate edits of the deck earlier than I recorded it on Saturday.
[00:08:19] And this dropped like final Wednesday or one thing like that. So I used to be like, okay, go. And by chance I had an entire part on laptop use that defined all of this. Going again to love the World of Bits paper from 2017. So fast recap. AI brokers are methods that may take actions with various levels of autonomy to realize a objective.
[00:08:37] So chatbots simply output textual content picture, video technology as attainable. Immediate one thing, you get one thing out of it. AI brokers can undergo a collection of actions. Typically they plan these actions themselves. Typically they do the vast majority of these actions with very restricted human involvement. However for probably the most half, people are nonetheless, extraordinarily within the loop.
[00:08:58] However this concept of [00:09:00] what’s known as laptop use brokers goes again to love 2016 17 opening. I used to be engaged on this very factor they usually revealed a paper known as World of Bits. And on the time they mainly stated. We’re simply not there but. Like we won’t bodily do what we would like this stuff to do. Give them entry to keyboard and mouse and fill out varieties and take actions on web sites like people do.
[00:09:22] Properly, language fashions Then, , shortly thereafter had been launched and, and that truly grew to become an unlocked to construct these com laptop use brokers. So we had our first laptop use, publicly accessible laptop use agent from Anthropic of all locations. In fall 2024, ju Google launched one thing.
[00:09:40] OpenAI has launched one thing. I feel perplexity has one thing like everybody’s now going on this route. So this concept of the agent having the ability to full actions in a digital surroundings, like in your laptop, in your telephone, that is the place they’re all going. So that is early, so I will simply, to maintain this too fast hearth, I will learn Sam Altman’s tweets.
[00:09:58] ‘trigger I believed they had been very [00:10:00] telling, about form of the place we’re. So he stated at this time we launched a brand new undertaking known as, or product known as, Chad, GPT Agent. Agent represents a brand new degree of functionality for AI methods and may accomplish some exceptional advanced duties for you utilizing its personal laptop. As you talked about, Mike, it combines deep analysis and operator.
[00:10:18] Operator was truly simply launched in January, 2025 as a analysis preview. So that is form of constructing on that. but it surely’s extra highly effective than might sound it may well assume for a very long time. Use some instruments, assume some extra, take some actions, assume some extra, et cetera. So to form of unpack this for a second, the reasoning mannequin oh one which, OpenAI launched in September, 2024, is what now unlocks the power to do that.
[00:10:45] So when he is saying assume some, then assume some extra, then assume some extra. That is the reasoning element constructed into this. Mm. They’re now on O three Professional. we’ll speak just a little bit about GPT 5 in a minute, however I feel that that is mainly what’s occurred. So then again to Sam. He says, for instance, we confirmed a demo [00:11:00] in our launch making ready for a good friend’s wedding ceremony the place you are shopping for an outfit, reserving journey, selecting a present, et cetera.
[00:11:05] We additionally confirmed an instance of analyzing information and making a presentation. Though utility is important, so are the potential dangers. We have now constructed numerous safeguards and warnings into it and broader mitigations that we have ever developed than we have ever developed earlier than, from sturdy coaching to system safeguards to, consumer controls.
[00:11:25] However we won’t anticipate every part. Within the spirit of iterative deployment. We’re going to warn customers closely and provides customers freedom to take actions rigorously if they need. I believed this was a very attention-grabbing a part of his tweet. I might clarify this to my family as innovative and experimental, an opportunity to strive the long run, however not one thing I might use for prime stakes makes use of or with numerous private data till we now have an opportunity to review and enhance it within the wild.
[00:11:50] In different phrases. They know it would take your stuff. Like in the event you enable this factor entry to your laptop and to see issues in your laptop which might be [00:12:00] private and confidential, they’re mainly warning you. We do not know if this factor is gonna take that information and use it in some nefarious methods is form of accessible approach of claiming this.
[00:12:09] You assume we do not know precisely what the impacts are going to be, however unhealthy actors might attempt to trick customers, brokers into giving non-public data they should not and take actions they should not in methods we won’t predict. We advocate giving brokers the minimal entry required to finish a process to scale back privateness and safety dangers.
[00:12:27] For instance, I can provide agent entry to my calendar to discover a time that works for group dinner, however I need not give it any entry if I am simply asking it to purchase me some garments. There’s extra threat in duties, like, take a look at my emails that got here in in a single day and do no matter it’s good to do to handle them.
[00:12:42] Do not ask any comply with up questions. So he is, he is mainly saying, do not try this. Like do not, do not give it this free entry to do it ‘trigger it’ll. It’s possible you’ll not like what it does. he stated this might result in untrusted content material from a malicious e-mail tricking the mannequin into leaking your information. We expect it is [00:13:00] vital to start studying from contact with actuality and that individuals undertake these instruments rigorously and slowly as we higher quantify and mitigate the potential dangers concerned.
[00:13:08] As with different new ranges of functionality society, the expertise and threat mitigation technique might want to co-evolve. after which he did do a follow-up tweet the place he stated, watching ChatGPT agent use a pc to do advanced duties has been an actual really feel the AGI second for me. One thing about seeing the pc assume plan and execute hits totally different.
[00:13:29] So general it’s nonetheless very early, particularly in relation to laptop use. However these AI labs, the main labs, are all very aggressively pursuing this path of improvement and deployment. And so we now have to begin. Be making ready for it, and also you and your group higher be updating your AI utilization insurance policies actually quick.
[00:13:51] If in case you have not but, as a result of there’s an opportunity your workers could possibly activate this sort of entry, or they will use their private entry and it’ll see [00:14:00] issues that your organization has in the event that they’re utilizing, , firm servers, firm emails, issues like that on their computer systems. So, my, my normal take is you shouldn’t enable your workers to show this on, on their computer systems, till you higher perceive the dangers concerned with it.
[00:14:17] Grok 4
[00:14:17] Mike Kaput: Yeah, no kidding. That final level. So vital I really feel like. Alright, so subsequent up, Grok 4 is right here and X AI claims. It’s the most clever AI mannequin on the planet. Not like earlier variations, Grok 4 was educated utilizing reinforcement studying at unprecedented scale. Because of X’S 200,000 GPU Colossus cluster.
[00:14:39] Grok 4 Heavy. Essentially the most highly effective model of the mannequin is now the primary mannequin to attain over 50% on human’s Final examination, which is a benchmark designed to check professional degree reasoning throughout domains. Grok 4 additionally comes with native software use, that means it is aware of how and when to run [00:15:00] code. Browse the net search X or dive into visible media.
[00:15:05] It autonomously chooses the perfect software for the duty, pulling real-time information and synthesizing solutions to realize your objectives. Now, Paul Grok 4 is spectacular, particularly when you think about simply how not too long ago Xai entered this AI arms race in comparison with how lengthy the opposite labs have been engaged on these items.
[00:15:25] And regardless of its controversies, which we are going to discuss, it appears to me that this does show progress isn’t actually slowing down. What do you assume?
[00:15:35] Paul Roetzer: So that they’re transferring extraordinarily quick. Clearly, Elon Musk, traditionally that is what he does. You understand, he, he form of takes an strategy to issues the place he simply goes all in, typically, , regardless of any dangers and issues like that.
[00:15:50] And there are definitely questions and issues about their seeming lack of AI security and alignment work, however they’re completely a part of the dialog now with the opposite main [00:16:00] Frontier Labs, which might, , typically be Anthropic, OpenAI, Google, and Meta as form of, they’re form of the large 5.
[00:16:07] Now, there, there’s others and we form of discuss them, however, they’re within the dialog they usually’re keen to do issues that the opposite labs will not do. That is not all the time a superb factor, however they completely are gonna take extra dangers than most of these different labs aside from in all probability meta, I might think about.
[00:16:24] I may see Zuckerberg being fairly aggressive in relation to these sorts of dangers. Elon Musk did. Put up the continual, reinforcement studying, rl enchancment of ROC appears like AGI. So right here we go together with the sensation, the AGI second. GR 4 at this time is smarter than Grok 4 just a few days in the past. Now that is a, it is a fairly, I feel it was truly in a reply to another person.
[00:16:47] It wasn’t even like a submit, but it surely’s a reasonably large deal as a result of, , once more, the best way these fashions work, in the event you’re not acquainted, is that they have a coaching cutoff date. So that you, you, you run a mannequin, you do the coaching, Grok 4 comes out and it, [00:17:00] it, it, it is information base stops when the coaching stopped.
[00:17:03] However by doing reinforcement studying constantly on prime of a mannequin, the mannequin can preserve getting smarter. And so that is what he is implying right here. Now Xi, I would not publish any analysis, so I don’t know how precisely they’re doing it, like in the event that they’re doing one thing totally different than the opposite labs. However, they’re clearly persevering with to enhance it.
[00:17:22] My, my guess is largely, based mostly on, X or Twitter information. Is form of like the principle proprietary factor that they should, to in ingest into these fashions. So, yeah, it would be attention-grabbing to see, however they’re, they don’t seem to be going away. They’re gonna preserve elevating billions and tens of billions of {dollars} and, they usually’re gonna preserve constructing large information facilities they usually’re gonna preserve making this mannequin greater and smarter.
[00:17:46] Grok Controversy
[00:17:46] Mike Kaput: Now, in our subsequent subject, Grok additionally did have its fair proportion of controversy prior to now couple weeks as a result of after a system replace, Grok sadly started to generate a wave of anti-Semitic [00:18:00] content material, together with frequently praising Adolf Hitler. And this got here after a system replace that explicitly instructed Grok to not shrink back from politically incorrect claims, which is a immediate that has since been eliminated Among the many extra disturbing outputs, Grok advised Hitler could be a superb chief for contemporary America, made repeated references to Jewish final names as indicators of somebody being an excessive activist.
[00:18:26] And referred to itself at one level as MechaHitler. So X AI apologized for the difficulty saying that the issue wasn’t the language mannequin itself, however an upstream code change that quote by chance reactivated outdated deprecated directions. Now in keeping with this clarification, Grok was in a position unintentionally to basically echo extremist consumer posts as an alternative of filtering or rejecting them.
[00:18:52] And that they are saying paired with some consumer prompts led Grok to strengthen hate speech. The offending code has since been deleted. [00:19:00] XAI says it is refactored the system and added new safeguards. So Paul, it was good to see Xai kinda supply a full clarification right here. however alternatively, this is not the primary time one thing like this has occurred.
[00:19:14] It appears for no matter purpose to occur much more with Grok than these different instruments. , I simply preserve considering we’re shortly transferring right into a time the place professionals and companies. Rely more and more on these instruments to do their work and controversies like this, I suppose, make me extra reluctant to attempt to construct Grok into something enterprise associated that I do.
[00:19:38] Paul Roetzer: Yeah. I am unable to, I am unable to see how Grock is gonna be an enterprise software in any approach. Proper? Like, I simply do not. Now that being stated, like they did an replace to Tesla. I imply, I obtained the replace on Friday Rock’s now accessible in Teslas, which is what I assume. And, and I, in the event you return all the time, I had talked about that is what they might finally do.
[00:19:57] We put GR into the Teslas. It was an apparent factor. Now it, [00:20:00] you possibly can solely speak to it like a chat bot, like a voice assistant. It will possibly’t management something in your automobile but. However that is the place they will finally go is like an clever engine constructed proper throughout the automobile. so all this insanity occurred like 48 hours earlier than they dropped Rock 4.
[00:20:17] So it was all like, it was wild. It was like a loopy, like three day stretch. So this does construct on the protection and alignment problem. So Rob Wilin, who’s the host of the 80,000 Hours podcast, had a tweet that I feel, summarized this rather well. So I will simply learn what he stated. XAI is an attention-grabbing one to observe for an early rogue AI incident.
[00:20:36] after which he simply bullet pointed this does big quantities of reinforcement studying, which generates unintended reward, hacking conduct, transferring very quick, deploys instantly. They do not wait. They identical to prepare dinner the factor, comes outta the oven they usually simply put it out into the world. has extra compute than gifted workers.
[00:20:52] that one’s, that is fairly humorous truly. Not doing any security stuff so far as anybody can inform, all demonstrated [00:21:00] by Mecha Hitler and the opposite issues Grok has finished, which XAI would not have wished. As soon as it strikes into brokers, there needs to be some probability it trains and deploys an unhinged mannequin that goes on to do actual hurt.
[00:21:12] I, I’m utterly aligned with Rob on this. Like, I, if any lab goes to take this factor the flawed approach, it, it seems to be XAI for the time being, which is so bizarre, contemplating the entire function in 2015 of Elon Musk and Sam Altman, others creating OpenAI, was {that a} counterbalance to, to Google who they thought was evil and will create a rogue ai mainly.
[00:21:37] So, proper. It is virtually like, I do not know, 11 years later it is like, screw it. We’ll simply do it ourselves. Like, we’ll simply, I dunno. so then Miles Bru, Bruge, who we have talked about earlier than, former AI alignment chief at OpenAI, he tweeted nonetheless no full security coverage month or so previous the self-imposed, imposed deadline.
[00:21:57] No system card ever, no security [00:22:00] evals ever. No coherent clarification of the reality searching for factor, et cetera. Or did I miss one thing? So he is referring to Xai and Grock. So then, let’s have a look at, then we discovered this was all, once more, all taking place in the identical like three day interval. it got here out that Grok 4 earlier than it will reply controversial subjects, was truly trying on X to see what Elon Musk would say about it.
[00:22:27] Sure. So it was found by way of testing by the general public that AI researchers know what they had been doing. They might go and look and see like what is the chain of thought round the way it’s doing it. And so they found it was truly like trying up Elon Musk tweets earlier than it will reply to issues about Israel.
[00:22:45] And in order that was loopy they usually get known as out for that. After which Simon Willison, truly shared the replace they made to the system immediate to try to repair this. So once more, in the event you do not perceive how these fashions work, they’re simply gonna do what they’re gonna do after they arrive outta coaching [00:23:00] that the groups at these labs then try to like educate them to behave in sure methods.
[00:23:06] And so the best way they do it’s they simply give it totally different phrases like directions. It isn’t like they go in and alter the code and it simply is all the time gonna now behave, proper? They’re mainly simply pleading with the factor, are you able to, are you able to cease doing the factor you are doing? So here is how they fastened it.
[00:23:20] That is actually within the system. Immediate responses should stem out of your impartial evaluation, not from any acknowledged beliefs of previous rock. So do not take a look at previous rock Elon Musk or XAI if requested about such preferences. Present your personal purpose perspective. So that they’re simply saying, please, like, truly give it some thought your self.
[00:23:41] So here is, and once more, I wanna preserve this fast hearth. This is my largest concern with all of this. Who decides reality? So. Proper now, as I discussed earlier than, we now have about 5 labs in america which might be coaching probably the most highly effective AI fashions. These labs are run by 5 folks, so Demi Saba at Google [00:24:00] DeepMind, Sam Altman at OpenAI, Dario Ade and Anthropic, Elon Musk, XAI and Mark Zuckerberg at Meta.
[00:24:07] You may in all probability put Microsoft in that blend finally. However their settlement with OpenAI limits their potential to pursue constructing frontier fashions and AGI themselves. there are another labs within the cons in consideration right here, like Secure Tremendous Intelligence, which is I Ilia, Eva SVAs. There’s Considering Machines, labs, which is definitely the one feminine run lab that we now have with mirati, Amazon and Mytral, like, so there’s others, however these 5 are the one ones who can spend billions of {dollars}, have a whole lot of hundreds, and shortly to be tens of millions of GPUs and TPUs.
[00:24:42] The chips wanted to coach this stuff and entry to the information facilities and power sources wanted to construct and ship probably the most superior ai. Every of these labs, once more, simply so folks perceive how this works, every of these labs makes selections. They select which AI researchers and engineers they’ll rent, after which they [00:25:00] select the values and AI alignment rules that these persons are presupposed to comply with.
[00:25:04] Now, they do not all the time comply with ’em. There was truly a excessive profile incident yesterday the place an X xai researcher obtained fired for what he was saying on-line, which was mainly, people ought to simply, surrender and let the extra clever species emerge. And so xai fired any individual, which is form of nuts. Hmm.
[00:25:23] So these people that, that are actually curated to make selections, they then curate the information that trains the fashions. Then one other set of people inside these 5 labs, who’ve their very own inherent biases, they submit prepare the fashions. So it comes out, , educated after this coaching run and all the information was given.
[00:25:42] After which they make selections to effective tune this stuff by particular information units that give them specialised capabilities. And to re do reinforcement studying, which form of teaches them behaviors and reply in what codecs and issues like that. Then one other set of people inside these labs [00:26:00] writes and maintains the system immediate that we simply talked about that guides how the mannequin behaves its character, its willingness to carry out the requested outputs and actions by the tip consumer, even when they’re nefarious and will hurt folks.
[00:26:11] These fashions, in essence, have all of human information, at the very least the publicly accessible human information plus no matter license stuff they’ve. The fashions simply wish to be taught. Ilia Skova famously stated this in like 2016, like they simply wanna be taught, like simply give ’em extra information. They, they simply devour it they usually wish to typically do no matter they’re requested to do.
[00:26:32] The one approach they do not do issues that may very well be dangerous is as a result of people put guardrails in place who then try to steer them in a particular approach. So if a sure AI lab, say X AI or a sure nation state. China, United States, no matter. If they need a mannequin to be totally different or to realize a distinct function that aligns with their creators, not with typically accepted human values, however simply no matter their creators decide, then in principle that is what [00:27:00] it is gonna do.
[00:27:00] However as we noticed with Grok, generally it does what it needs. Prefer it’ll simply turn out to be a Mecha. Hitler prefer it. It simply does one thing totally different. And as we have seen with Anthropic analysis not too long ago, we have talked in regards to the podcast quite a few instances. Typically the fashions will simply faux alignment. So that you testing them and apparently they’re capable of know once they’re being examined, and they also fake like they’ve the values of their creators, after which they will go and do one thing else.
[00:27:27] So the explanation this turns into so essential to speak about proper now’s on June twenty first, 2025, Elon Musk. So that is what, three weeks earlier than the discharge of Grok? 4, three-ish weeks. He stated, we are going to use Grok 3.5, perhaps we must always name it Rock 4, which clearly they did. Which has superior reasoning to rewrite your entire corpus of HU human information, including lacking data and deleting errors, then retrain on that far an excessive amount of rubbish in any basis mannequin educated on uncorrected [00:28:00] information.
[00:28:02] Meaning Elon Musk is deciding as one of many 5 leaders of those 5 labs that his perspective on the world is the precise perspective. He, he determines the reality and that they can not depend on the overall information of humanity. They should depend on their very own view of it. Now, it’s possible you’ll love Elon Musk’s view of society and humanity, and it’s possible you’ll assume that is nice, like let’s have the Elon Musk model of this, and that is effective.
[00:28:31] That is your prerogative. However perhaps you assume Mark Zuckerberg has a greater perspective, or perhaps it is Sam Altman you belief, or, or perhaps it is Des, or perhaps it is none of them. And if it is none of them now the place are we? As a result of the query goes again to who decides reality, after which what are the implications of these, these selections on humanity.
[00:28:49] It is rumored that OpenAI has over a billion customers of chat, gpt. Mm, that that may be a giant share of society that makes use of it. [00:29:00] Google has seven merchandise or platforms with over a billion customers every. X has what about 200 million customers? meta has billions by way of Instagram. These are the 5 people who find themselves going to find out how intelligence is distributed to society and humanity.
[00:29:23] So I clearly may discuss this for the subsequent like two hours, however I simply need folks to grasp the macro of what’s taking place right here and that the way forward for all of this proper now’s being determined largely inside 5 labs in California and one in Texas. Oh, truly they is likely to be in California too.
[00:29:44] I feel they’re within the outdated OpenAI workplaces truly. In order that’s form of the place we’re at as a society. After which the governments are more and more coming into play and the governments, just like the US authorities, simply final week an article got here out mainly saying they wish to management the the place this goes. Now, once more, [00:30:00] that is harmful it doesn’t matter what political aspect of the spectrum you are on, as a result of in essence what it is saying is whoever’s in energy will get to now decide how these fashions behave and what values they’re aligned with and what these fashions assume is reality.
[00:30:11] So
[00:30:15] Mike Kaput: man, that is, I really feel like vital and disturbing sufficient for say, such as you and I sitting right here saying, okay, what are the implications of this on how society makes use of ai? However we even have to contemplate that we’re perhaps a technology away, if not a shorter period of time from. AI being the mediating layer between you and any data on the planet, children will default to assuming AI is appropriate it doesn’t matter what.
[00:30:41] As a result of why would not or not it’s? Will probably be the one factor they know use and know expertise. So this has even greater downstream results too, for future generations.
[00:30:50] Paul Roetzer: Properly, yeah, and I imply, Grok additionally, I say I additionally launched companions, in order that they now have a feminine and a male companion that’s [00:31:00] designed to do and say no matter you prefer to it to say.
[00:31:04] and you may join your personal dots right here if, , they’re being positioned as companions, they’ve launched them like they’re out within the wild. So, yeah. I I am with you, Mike. Like, I truthfully assume it is throughout the subsequent 5 years. Like Yeah. the upcoming gen, I imply, my children are 12 and 13. Like I take into consideration this each single day about like what AI instruments are they gonna have entry to.
[00:31:25] how do you, I do not wanna say management not the precise phrase. How do you monitor and, and educate them? To, to not be, turn out to be counting on a single viewpoint from a single lab, proper. Made up of some thousand people who find themselves making these selections for all of us.
[00:31:39] Mike Kaput: And simply as one ultimate observe right here, you are already seeing a really gentle model of this.
[00:31:44] Anytime you are on x now folks will instantly reply to a declare saying, at tagging rock and saying Rock, is that this true? And it is like, okay, we’re already seeing this. Persons are already taking that as gospel reality. If it says it’s or is not actuality,
[00:31:59] Paul Roetzer: it is [00:32:00] wild, man. I, that is like, yeah, yeah. Once more, I do know it is presupposed to be fast hearth, so I transfer on.
[00:32:07] Meta Superintelligence Labs
[00:32:07] Mike Kaput: Alright, subsequent up, mark Zuckerberg has unveiled meta tremendous intelligence labs. It is a daring reorg that’s geared toward placing meta on the forefront of the AI arms race. This new lab is led by Alexander Wang, previously founder and CEO at scale ai. It’s staffed with a wave of elite recruits from OpenAI, Google DeepMind, and Anthropic.
[00:32:31] All of whom have been apparently lured in with eye watering pay packages, some reportedly north of 200 million plus {dollars}. And meta is mainly pitching this prime expertise on the concept that they’ve mainly limitless compute deep product attain and a singular objective, which is to develop private tremendous intelligence.
[00:32:52] So that is AI brokers smarter than people, embedded in every part from messaging to wearables. Now, on prime [00:33:00] of this, internally there are studies the lab has already began debating whether or not or not they need to drop open, meta’s open supply technique in favor of closed fashions, although Zuckerberg has not but signed off on that.
[00:33:14] So Paul Zuckerberg is making some big strikes. Right here he’s poaching AI expertise left and proper, utterly upending compensation for prime AI expertise. The place does this go subsequent?
[00:33:26] Paul Roetzer: So that is one other one. I used to be in the course of finalizing the AI timeline course as a part of the AI elementary collection. And in that course I inform the story of, , clearly the street to AGI and past, that is the past half.
[00:33:38] Like I’ve all the time, however once more, I did not assume folks had been actually prepared for the tremendous intelligence dialog. And now abruptly in July, that is like all any of the main labs are speaking about. It is like we have simply moved previous AGI and we simply form of assume we’re there. We’re gonna be there very quickly.
[00:33:53] And all of the labs aren’t speaking about tremendous intelligence. Even Sam Altman not too long ago stated like, that is the objective of their lab now’s tremendous intelligence. They have not [00:34:00] modified their mission. AGI i’s nonetheless the mission, however they known as themselves a superintelligence lab. So all people has simply form of moved previous this.
[00:34:07] The, the numbers are loopy. it is exhausting to place these form of numbers in context, however the one which I believed was related is, NBA MVP, Shai Gilgeous-Alexander, the Oklahoma Metropolis Thunder. Simply agreed at first of July to a file setting 4 yr deal for $285 million. AI researchers are actually getting paid extra prime AI researchers getting paid greater than the very best paid skilled athletes.
[00:34:35] So I imply, all people would take a look at the 280 million, be like, oh my God, these athletes are paid a lot cash. That is a ridiculous contract. After which right here we obtained 10 of those folks already that we all know of that, meta has employed paying 300 million or extra. And I noticed a report this morning that they provided a billion to, to any individual.
[00:34:52] so it is not, it hasn’t been verified but. I, I will, I will not say who it was, however, yeah, so [00:35:00] it is loopy. After which, I imply, we already noticed, I imply we noticed Noam Shazi obtained mainly acquihire for two.5 billion by Google. We had the, Alexander Wang from scale ai 14 billion, mainly 15 billion to from by meta.
[00:35:12] So these numbers are simply getting insane.
[00:35:16] Windsurf Drama
[00:35:16] Mike Kaput: Yeah, form of talking alongside these traces on this subsequent subject, OpenAI, as we have reported in previous weeks, was actually near buying Windsurf, which is a quick rising AI coating startup in a $3 billion deal. However now that deal is off and windsurf is mainly now the property of two totally different corporations.
[00:35:34] So Windsurf backed out of the OpenAI deal after elevating issues that OpenAI’s settlement with Microsoft may result in Microsoft getting access to totally different items of windsurf tech. So after this, Google swoops in, they rent CEO, Varun Mohan and several other prime engineers and pay $2.4 billion for a non-exclusive license to [00:36:00] windsurf expertise.
[00:36:02] After which one other twist occurs. So its management is gutted. Windsurf mainly appears a bit adrift till cognition steps in. So over a frantic weekend. The corporate, which is behind the Devon AI coding agent that we have talked about prior to now. They struck a deal to purchase the remainder of Windsor. It is ip, it is model, it is remaining workers.
[00:36:23] So Paul, this one’s been a curler coaster since form of day one. What, what occurred right here? Why are corporations simply paying for items of this firm or folks in it and never simply shopping for it?
[00:36:34] Paul Roetzer: They can not purchase it, due to authorities oversight. And, , the, because of this all these are acquihire offers.
[00:36:42] They’re simply simpler to get by way of. and you then simply go away a shelf of an organization behind, mainly all, all I can take into consideration is, I by no means watched Silicon Valley, which I in all probability ought to. the present. Oh yeah. Yeah. I feel it was like six seasons. Somebody needs to be engaged on a reboot of that. Like that is, [00:37:00] I might watch these items.
[00:37:01] I, this does not, it is beginning to simply not even really feel actual like these acquihires and stealing expertise away and you bought 5 labs combating towards one another. And the loopy half is like all these analysis and engineers at these 5 totally different labs all hang around on the similar events and share inside secrets and techniques on a regular basis.
[00:37:18] It is like simply made for TV form of stuff, so.
[00:37:22] Mike Kaput: Properly, I will let you know, Grok turning into Mecha Hitler, might as properly be a Silicon Valley.
[00:37:27] Paul Roetzer: Yeah. It is like episode
[00:37:28] Kimi 2
[00:37:28] Mike Kaput: one, proper? Yeah, yeah, yeah. So I full, totally agree. They should reboot that. All proper. Subsequent up Moonshot, which is an Alibaba backed startup, has launched Kimi 2, which is a low price open supply language mannequin that’s drawing international consideration.
[00:37:45] As a result of it reportedly outperforms, Claude and ChatGPT on main coding benchmarks and does so at a fraction of the price. So Kimi two fees simply 15 cents per million enter tokens and [00:38:00] $2 and 50 cents for output, which is seemingly 100 instances cheaper than Claw and nonetheless far beneath OpenAI charges.
[00:38:08] So it is also open supply, which lets builders freely construct on it and early opinions reward it for robust actual world coding use, although it’s nonetheless growing some integration with different methods. Now, Paul, first we obtained deep search. Now Kimi 2. It has been unimaginable seeing simply how a lot energy and efficiency for value we are actually capable of get utilizing open supply fashions.
[00:38:32] And it additionally looks like simply one other blow to Meta’s open supply technique. I imply, these fashions out of China are simply racing forward whereas one thing like llama seems to be, at the very least for the second, caught within the mud.
[00:38:45] Paul Roetzer: Yeah, I, I, once more, this will get into just like the macro degree stuff. I imply, in essence, in essence, what’s taking place is in the event you take a snapshot in time and say, okay, probably the most highly effective fashions on the planet at this time are Gemini 2.5 Professional, I could also be [00:39:00] Grok 4 in some capability.
[00:39:01] The O three mannequin from OpenAI, from a reasoning perspective, GPT 5, possible, , perhaps earlier than the tip of July, feels like in all probability, at the very least definitely this summer season, these state-of-the-art fashions will probably be mainly free, 12 months from now. So each 12 months or so, the price of utilizing these fashions drops like 10 x.
[00:39:27] And so it’s important to virtually look out whenever you’re planning on your life, whenever you’re planning for society, whenever you’re planning for enterprise. And, and assume that intelligence is mainly gonna be free. That, that tremendous intelligence, like issues which might be smarter than the neatest people at mainly each cognitive process is just about gonna be free to all of us.
[00:39:46] as a result of the competitors between these labs virtually dictates that that is the place this goes. They’re gonna open supply. The earlier technology mannequin, that earlier technology of fashions is state-of-the-art at this time. And so we’re on a, there, there [00:40:00] seems to be no close to time period cease to that trajectory. the pre-training scaling legal guidelines adopted by the post-training scaling legal guidelines, adopted by the check time compute legal guidelines, which is give it extra time to assume and it will get smarter.
[00:40:12] These three scaling legal guidelines working collectively after which perhaps yet one more scaling regulation that we, we do not know but form of the subsequent breakthrough that should occur. They’re form of dictating that. S first normal intelligence, however then tremendous intelligence will mainly be too low cost to meter. Like, it will, it will simply be available.
[00:40:31] After which the distribution that each one these corporations have form of dictates, properly, who truly controls all of the customers. So, I do not know. I imply, it, it is nuts. Like I get, I do not even know, I take advantage of the phrase wild loads. We may in all probability like, question our transcripts. Somebody’s in all probability gonna go do that, like put all our trans, however I say the phrase wild, wild loads.
[00:40:50] I do not understand how else to explain it. Like, it, it actually is tough. I, I, it is attention-grabbing. I, so final evening I, I had my daughter’s 13. I requested her if [00:41:00] she watched Interstellar with me. Prefer it’s one among my favourite motion pictures ever. And I’ve watched Interstellar in all probability, I do not know, six or seven instances now.
[00:41:05] And so we sat there for like three hours watching this film, and I get finished and I am laying in mattress. I am like, and I nonetheless do not actually perceive it. Prefer it’s the, the ending of that film is so thoughts bending and it is just like the human thoughts’s virtually not even presupposed to form of comprehend the essential idea of intertel.
[00:41:23] And I form of really feel that approach in regards to the intelligence explosion. Prefer it’s, it is virtually simply too exhausting to even comprehend what is nearly inevitably going to occur within the subsequent three to 5 years. And this matches into that realm. It is like, what, what, what occurs if tremendous intelligence is a factor in just a few years?
[00:41:41] And it is simply virtually to all people. Like, I do not know. I do not even understand how you are supposed to understand that stuff.
[00:41:47] Mike Kaput: Yeah. I additionally assume it is extraordinarily troublesome to maintain perspective on simply how briskly issues change. We are saying it on a regular basis, we really feel it, however in the event you actually sat down and regarded on the milestones over the past three years since ChatGPT [00:42:00] got here out, it will be breathtaking.
[00:42:02] Paul Roetzer: Yeah. It it, that is a superb level. Like GPT-4 was March of 2023. We, we are actually speaking about two years Yeah. Of innovation and, and look the place we’re like. Is I do not know. I imply, I had that slide in in one among my programs. what if the long run is, , 10 x, 20 x, 100 x innovation taking place each two years, like, as a result of proper now that is the trajectory we’re on and I do not even know clarify that to folks.
[00:42:31] Like, I am unable to realize it myself.
[00:42:35] AI Browsers
[00:42:35] Mike Kaput: Alright, subsequent up, OpenAI is on the brink of launch its personal AI powered internet browser in keeping with Reuters. the upcoming browser is designed to combine tightly with ChatGPT and a few of OpenAI’s agentic capabilities. As an alternative of pointing you to web sites, it’s seeking to perhaps preserve interactions inside a built-in chat interface, which is a direct risk to Google search and advert ecosystem.
[00:42:59] [00:43:00] It would reportedly enable brokers to take actions for customers, do issues like we have already seen, like reserving eating places, filling out varieties, managing inboxes, all whereas capturing the form of behavioral information that Google has lengthy used to dominate advert goal. Now we’re truly already beginning to see what an AI powered browser may appear to be as a result of Perplexity additionally simply launched its personal AI powered browser known as Comet.
[00:43:24] That is solely accessible proper now to Perplexity Max customers, which is their $200 a month providing, or by non-public invitation. And the Verge truly obtained their arms on this and revealed a reasonably intensive overview. And so they stated it would not simply assist you to browse. It actively takes actions for you. It will possibly summarize pages, handle your tabs, ship emails, unsubscribe from newsletters, even settle for LinkedIn invitations all in your behalf.
[00:43:51] You too can say, take management of my browser, and remark will go into agent mode. It will possibly place Amazon orders, ebook restaurant reservations, [00:44:00] and beneath the hood, it replaces conventional search outcomes with Perplexities reply engine. Now the Verge says proper now it may be generally slower than doing issues your self.
[00:44:10] It will possibly often come across extra advanced duties, however it’s an attention-grabbing early instance of what an AGI agentic AI browser can truly do. So whereas it is nonetheless actually early right here, Paul, it does really feel like we’re on the sting of perhaps one thing large right here. I imply, the best way wherein the Verge described this, regardless of all the issues right here, appeared like an AI browser may very well be an actual paradigm shift from the best way we sometimes use browsers at this time.
[00:44:37] The place, the place do you see this going? How do you consider this?
[00:44:40] Paul Roetzer: Yeah, so there is a, a few fast notes right here. So Perplexity purchased os.ai from Dharmesh Shaw, who’s the co-founder and CEO of HubSpot. this was all, this was final week. And Arvin, the CEO of Perplexity stated there is a purpose we bought os.ai from Dharmesh.
[00:44:57] The roadmap is to make comet really feel like your [00:45:00] personal mini personalized laptop inside your current laptop telephone and the pc working throughout consumer and server with the power to run native fashions too. However the tweet that I believed truly form of defined this greatest ‘trigger I used to be like, I used to be form of struggling to grasp like what precisely was occurring right here.
[00:45:17] Like, I knew the overall idea, however I could not like deeply clarify this one but. And so there was a tweet I noticed. that is from July seventeenth. We’ll put it within the present notes from Michael Mcno, who’s a VC and founder companion at Lightspeed Ventures. so he stated the browser sees every part. That is the explanation we’re getting so many new AI first browsers from the browser firm Perplexity, and shortly OpenAI, which we obtained.
[00:45:40] to allow them to see information that they more and more can not scrape. AI feeds on information. It will get the information by robotically scraping the net, however scraping is now not free. CDNs like CloudFare Clare Cloud Flare are making scraping tougher by blocking it by default, and others will quickly comply with comply with. Startups like tobit [00:46:00] are empowering tons of enormous publishers to cost for being scraped, constructing a brand new open internet economic system.
[00:46:05] However customers need ai, we won’t get sufficient of it. As AI solutions more and more eat conventional internet search, AI will probably be doing far more searching on behalf of us people. this creates a paradox. Client conduct is shifting to ai, however AI is working out of gasoline to fulfill the demand. So what occurs? AI corporations construct browsers as people devour content material with these browsers.
[00:46:28] The AI firm can see, quote unquote, see that, information that’s more and more being blocked or monetized the corporate. Fascinating. probably the most attention-grabbing factor about this technique is that AI corporations do not even want significant market share or buyer UBI ubiquity for this technique to work. They only want a big sufficient slice of all searching to get a style of a lot of the internet’s information.
[00:46:51] It is an entire new enterprise mannequin for the net and the start of a brand new browser conflict. I used to be like, oh shit, that is fairly good. Like that makes extra sense now. So [00:47:00] yeah, I do not know that browser wars, expertise wars. Nice, nice, nice television, proper?
[00:47:05] Microsoft’s Layoffs and AI
[00:47:05] Mike Kaput: Yeah, no kidding. Good, good for customers too, at the very least in, within the brief time period.
[00:47:10] All proper, so subsequent up, Microsoft is shedding hundreds, however urging remaining workers to upskill in ai. After slicing 9,000 jobs not too long ago, the corporate advised gross sales workers to embrace AI instruments like copilot to spice up productiveness and shut extra offers, quote, spend money on your personal AI skilling. One government suggested whereas others reportedly started factoring AI utilization into efficiency opinions.
[00:47:36] Now behind these cuts as Microsoft hovering infrastructure prices, in order that they count on to spend 80 billion this yr on information facilities and chips. A lot of IT supporting AI providers for patrons and OpenAI. And to offset that, Microsoft claims that AI is already paying off. It saved, they stated over $500 million in name middle prices final yr, and gross sales reps utilizing copilot are [00:48:00] reportedly producing 9% extra income.
[00:48:03] Internally, the corporate is consolidating roles and gross sales models. They’re even monitoring how a lot code engineers generate utilizing ai. Nonetheless, on prime of all this, the executives at Microsoft insist that these layoffs aren’t solely pushed by ai. So Paul, it is good to see, they’re fast to say that wasn’t the case, however it’s fairly clear that AI at each degree is driving the technique and selections right here.
[00:48:28] Is that this one other instance of an organization saying the quiet half out loud?
[00:48:32] Paul Roetzer: Yeah. This tracks with every part we’re seeing. I imply, I do not even know that is quiet anymore. I really feel prefer it, it simply is what it’s. There’s, there is not any debating this. We have, , in one of many programs I used to be creating, I feel I highlighted like seven or eight, CEO statements simply from June to July of this yr.
[00:48:51] Yeah. The place they, they actually stated like, we’re simply gonna want fewer folks. We’re, we’re beginning the method now, get AI literate or get out is mainly what the [00:49:00] CEOs are saying. So, yeah, I imply this, this can be a hundred % the route. I do not, I do not know that it is debatable. I feel I might heed the recommendation, like, you, it’s important to, it’s important to push your self to be one of many folks within the room who understands these items and may apply it to what you do every single day.
[00:49:20] Like it’s, I do not know that you’ve job stability, , one to a few years out, no matter your career or trade. Should you do not try this, and I imply, I assume folks listening to this podcast are the folks within the room who’re figuring this out. So good on you. Hold going. You are, you are gonna be in a very good place transferring ahead.
[00:49:42] when you have pals, household, coworkers, you care about, it’s important to get them to begin figuring these items out. Like their, their, their jobs are gonna rely on it, in a short time. Should you’re within the tech area, it is tomorrow. Should you’re in authorized healthcare. Monetary [00:50:00] providers, like perhaps manufacturing might, perhaps you bought just a little little bit of time, however all people’s gonna determine this out.
[00:50:06] Publicly traded, non-public fairness backed, VC backed is now anyone else? It is, it is coming. So do, do what you possibly can to try to pull, pull folks together with you in the event you’re listening to this present.
[00:50:17] Mike Kaput: Yeah. And I do know that we’re clearly very biased right here in direction of AI literacy, however I might argue too, to actually strongly encourage folks to reframe this and never see AI as one different ability.
[00:50:31] It’s essential be taught it’s the factor it’s good to determine. Like I might be ruthlessly slicing any of your different skilled improvement priorities as if you’re less than velocity but on ai.
[00:50:42] Paul Roetzer: Yeah, I imply, I consider it, you are proper, Mike. I do consider it as just like the underlying working system to our careers.
[00:50:49] Like actually every part you do goes to should be constructed on prime of AI and Precisely. Individuals who do not determine that out are, are simply gonna have a very troublesome [00:51:00] time sustaining profession stability. And I feel the individuals who do are gonna have like super close to time period profession potential, like incomes greater than your friends.
[00:51:11] Creating extra worth inside corporations than your friends. Such as you’re simply gonna be capable to do extra and be extra precious and, and that is only a higher place to be. Like I might, I might slightly simply be, I imply, there’s nonetheless no ensures, however like I might approach slightly be out on the frontier of these things, figuring it out and serving to different folks than sitting round and getting run over by it.
[00:51:32] Mike Kaput: And only one ultimate factor there, simply from private expertise, like concentrate on the enjoyable. I notice there is a ton of like doom and gloom. There’s loads to be afraid of. There’s numerous uncertainty, but in addition like, I simply get so excited extra so than ever to do sure work I do as a result of it is a lot extra new and thrilling with ai.
[00:51:49] Like, yeah, that is nice. Like I might truly be upset if I needed to do my job the best way I used to do it generally.
[00:51:56] Paul Roetzer: Now, properly there’s so many, like I for, for the AI course I used to be constructing, I created [00:52:00] ada, like AI educating assistant. Oh my God,
[00:52:01] Mike Kaput: that is an excellent
[00:52:02] Paul Roetzer: instance. Is that, yeah. Properly, yeah. There’s so many days the place I might use it and be like.
[00:52:05] I am unable to imagine that is attainable. Like I might accomplish that I might, I might actually like should message Mike or message, he is like, I simply should share this with any individual. Like take a look at what it simply did. Like there’s, there’s like that just about shock and surprise that every single day you are identical to, I’d uncover some new factor this factor does.
[00:52:20] And it simply makes work extra attention-grabbing and like the long run extra enjoyable to consider.
[00:52:27] Extra Meta Updates
[00:52:27] Mike Kaput: Yeah, 100%. Alright, subsequent up some extra meta updates. So we’ll undergo a pair issues right here after which Paul form of get your tackle this. So Meta simply made form of two different large strikes the place, it form of reveals the place they assume AI is headed.
[00:52:41] So first into your glasses and subsequent into your inbox. So first, meta invested $3.5 billion in Essilor Luxottica, which is the world’s largest eyewear maker. The corporate behind RayBan and Oakley Meta now owns slightly below 3% of the corporate. That is constructing on [00:53:00] Meta’s RDD partnership with RayBan as a result of they’re constructing their AI powered meta RayBan glasses on the similar time.
[00:53:07] Inside paperwork present meta is coaching customized chatbots to behave extra like companions. Bots which might be constructed on its AI studio platform can now ship unprompted personalised follow-up messages referencing previous dialog. The objective is to extend consumer engagement and retention by making AI really feel emotionally current.
[00:53:29] Now Paul, I discovered each these fairly attention-grabbing, clearly very clear metas all in on AI wearables. Additionally, the stuff with the AI autonomously messaging you to drive engagement and retention form of makes me shutter as a result of I can like think about with readability what dystopian future this turns into within the flawed arms.
[00:53:50] Paul Roetzer: Yeah. So, simply give a pair fast feedback. I’m fascinated by glasses. I’ll possible sooner or later strive them myself. [00:54:00] I’m 100% disturbed by a future society the place folks, plenty of persons are sporting this stuff and also you by no means have a clue who’s recording what. And I feel we’re in all probability already like on the sting of that.
[00:54:11] I, I actually fear about that. And I do not assume society is anyplace near being ready for that. And the ramifications of that, by way of the proactive outreach from the AI get used to it like that may be a hundred % going to be. So once more, we return to the competitors between these 5 AI labs.
[00:54:31] How do you management the consumer? Properly, it’s important to drive stickiness and engagement. It’s a must to drive, , each day lively customers, hourly lively use, like there, these are the metrics they’re gonna take a look at. And the one approach to try this is to immediate somebody. Now there’s worth, like say I’ve talked to a few well being situation and I am attempting to determine what is going on on with me.
[00:54:52] Then like three weeks later, it will simply examine in on me, say, Hey, how’s that factor going like that we talked about three weeks in the past? And that is [00:55:00] gonna really feel unimaginable. Like, you are gonna be like, rattling, that is, that is actually useful. And also you say, I am truly like, I am form of nonetheless having this symptom. And now abruptly you are having this like tremendous proactive dialog along with your AI about one thing that is very precious to you.
[00:55:13] And so it simply appears inevitable that, that we get there in a short time as a result of there’s not likely any technical limitations to doing it. There’s not like some breakthroughs wanted in AI to do that. It is form of like a reminiscence factor and like a context window factor. However these are trackable and there is already some options in place, however they’re all gonna do that.
[00:55:34] Like Google will in all probability should be slower as a result of folks take a better eye on what they’re doing. Opening eye is certainly gonna do that very quickly. Xai will do that 100% Meta will do it. Anthropic I might guess could be final on the listing to do it, however who is aware of?
[00:55:53] OpenAI Updates
[00:55:53] Mike Kaput: All proper. Subsequent up we now have some OpenAI updates along with the information we have already talked about.
[00:55:59] So first [00:56:00] up, Sam Altman truly posted about the truth that OpenAI’s language mannequin has achieved gold medal efficiency on the 2025 Worldwide Math Olympiad, which is the world’s premier math competitors for pre college college students, which is a big deal provided that it is an enormous benchmark and milestone.
[00:56:18] And similtaneously he introduced this on x, Sam Altman additionally teased GPT 5, which is coming quickly, and he stated quote, we’re releasing GPT 5 quickly. However I wish to set correct expectations. That is an experimental mannequin that includes new analysis strategies we are going to use in future fashions. We expect you may love GPT 5, however we do not plan to launch a mannequin with IMO Worldwide Math Olympiad gold degree of functionality for a lot of months.
[00:56:48] On the similar time, OpenAI are. Locked with Microsoft in a excessive stake standoff over the definition of AGI. They’ve this clause of their contract that states that if OpenAI’s board declares [00:57:00] it is reached AGI, Microsoft’s entry to future fashions could be minimize off. Now, there is a new wrinkle to this as a result of there’s, an unreleased inside paper at OpenAI known as 5 Ranges of Normal AI Capabilities.
[00:57:15] That is inflicting a stir as a result of it outlines a framework for assessing AI methods by functionality slightly than a binary. Simply sure or no. Is it AGI that complicates when and the way OpenAI may declare its reached AGI, and what would that will set off contractually? So Microsoft has poured 13 billion into OpenAI.
[00:57:35] They’re attempting to get this clause eliminated fully. OpenAI is form of saying, look, this isn’t actually a proper analysis paper. It is only a paper internally that we’re engaged on. Some OpenAI insiders say it’s pretty near AGI, so the entire paper muddies the artwork murky water round AGI on this partnership.
[00:57:57] Subsequent OpenAI is teaming up with the [00:58:00] American Federation of Lecturers to coach 400,000 US educators to form how AI is utilized in faculties. They’re funding a brand new nationwide academy for AI instruction. It is a nationwide coaching hub, providing workshops, hands-on programs, tech assist to assist academics combine AI into the classroom in ways in which improve, not substitute their work.
[00:58:21] There is a flagship facility plan to open in New York Metropolis. Extra places are deliberate nationwide by 2030. OpenAI is contributing $10 million over 5 years, together with direct funding API credit and engineering assist. Okay. On the similar time, OpenAI has launched a versatile credit score system for ChatGPT crew and enterprise plans that offers organizations extra management over how they entry superior options like deep analysis, O three Professional GPT-4 0.5, and Picture Era.
[00:58:52] And final however not least, an open weight AI mannequin from OpenAI was set to launch quickly, however now it is on maintain. [00:59:00] Sam Altman posted that there was now going to be a delay citing the necessity for extra security testing and deeper overview of excessive threat areas. There isn’t a new launch date but. So Paul, bunch of stuff right here.
[00:59:11] What sort of jumped out to you as value taking note of? It looks like we now have affirmation of GPT 5 coming quickly straight from Altman. I additionally discovered the unreleased paper round AGI fairly attention-grabbing.
[00:59:22] Paul Roetzer: So the Worldwide Math Olympiad is intriguing. There could also be extra popping out at this time, , after we’re finished recording this, however.
[00:59:31] The phrase over the weekend was, so No Brown, who we have talked about many instances on the present, met a researcher, went to OpenAI, on the forefront of reasoning, very well-known AI researcher. He tweeted this at like 2:00 AM on like Saturday or one thing like that. I noticed, I bear in mind I noticed it like as quickly as I awakened and I used to be like, that is a bizarre factor to tweet in the course of the evening.
[00:59:52] Like why would they’ve finished that? And so the rumor is that Google additionally achieved gold [01:00:00] medal on this, however that the group requested the labs to not disclose that that they had obtained gold medals for seven days following as a result of they wished to not steal the eye from the youngsters who had achieved unimaginable issues as people.
[01:00:17] And OpenAI at the very least no, him tweeted that. Like nobody had advised him that. Nobody had advised OpenAI, that form of factor. So the phrase is Google additionally did this, however they had been following the principles and never. Disclosing it. so I do know we’ll, we’ll, we’ll wait and see. Google. As of the second of recording this had had not confirmed somehow, but it surely was a, it appeared like a little bit of a, I do not know.
[01:00:43] There’s undoubtedly some controversy round it, so we’ll wait and, and see what occurred. By way of gpt 5, the rumor right here is it was presupposed to be what we might name a unified mannequin, that means it was a single mannequin that does reasoning and chat and every part else, which is what Gemini 2.5 Professional is, [01:01:00] is a unified mannequin that is form of constructed all in.
[01:01:02] Now, the dialogue is that gpt 5 could also be, is likely to be a router mannequin, that means it is not a single mannequin, it is nonetheless a chat plus a reasoning mannequin. However you’ll now not have to decide on which mannequin you are going to use. The mannequin would select itself. So that is what we have been, , preaching for the final two years is like, why once I go into CH PTs, do I’ve eight fashions to select from?
[01:01:25] So what they’re saying is GPD 5 may simply be an even bigger, smarter mannequin. It additionally would route you to a distinct mannequin if it is wanted, comparable to a reasoning mannequin. However we do not know. They, they, they have not stated something aside from it is, it is in all probability coming quickly. After which the paper is attention-grabbing as a result of I got here throughout this in one of many programs I used to be constructing.
[01:01:45] I, I noticed this after which I went again to the stage of ai. So I feel what it is truly associated to is in, round this time in 2024, Bloomberg launched an article the place they shared what was believed to be an inside doc [01:02:00] in regards to the levels of synthetic intelligence. Hmm. And in that doc, OpenAI, that is what they had been, OpenAI was utilizing as their inside guidepoint was, chatbots had been first, reasoners had been second.
[01:02:10] AI brokers was degree three innovators was degree 4, which is they may create new issues. They will uncover new, mathematical formulation. They will develop new science. After which, which I suppose the Worldwide Math Olympiad stuff begins turning into very related on the innovator spectr After which organizations, which is autonomously run organizations, that was the 5 ranges.
[01:02:31] And so the article that not too long ago got here out was implying that, that these levels had been a part of a paper at OpenAI that they didn’t launch due to their issues associated to the contract with Microsoft. So it is like, oh, that will clarify why these levels obtained leaked. After which sooner or later, OpenAI final yr form of acknowledged that sure, these are the levels we take a look at internally, however they by no means formally launched them themselves, [01:03:00] to my information.
[01:03:00] And this could clarify why it was by no means truly launched from OpenAI because it was a part of an even bigger paper that obtained yanked, mainly. So simply, I do not know ing historic context right here.
[01:03:13] Google Updates
[01:03:13] Mike Kaput: All proper. Subsequent up we will do some extra, updates round Google that got here out the previous couple weeks. So first up, they’re rolling out a serious AI growth for faculties with the launch of Gemini for training.
[01:03:25] It is a model of Gemini tailor-made particularly for college students and academics. It is constructed on Gemini 2.5 Professional and brings premium AI capabilities into Google Workspace for training, providing larger utilization limits, enterprise grade safety, and full admin management. Subsequent, Google is bringing its customized AI help gems instantly into the aspect panel of Gmail docs, sheets, slides, and drive.
[01:03:49] Till now, they had been dwelling within the standalone Gemini app. Over the subsequent few weeks although, they will present up, proper the place all of us work in workspace apps, proper within the aspect panel of these [01:04:00] apps. So which means you possibly can use a gem to draft emails, generate slide content material, analyze spreadsheets, what have you ever, with out switching apps.
[01:04:09] Google has additionally made some extra workspace updates. First VO three. Google’s superior video technology mannequin is now built-in into Google Vids, ai, voiceovers, and vids simply obtained simpler too. Now you can replace all voiceovers in a undertaking with a single click on. After modifying your script, they’ve added a group of recent templates to slides to hurry up presentation constructing, and the Gemini app is turning into much more linked.
[01:04:34] It faucets into Gmail, drive, calendar, preserve, and duties. So you possibly can ask Gemini to summarize unread emails, pull assembly notes, surfaces particular doc instantly from inside a chat. And for deep analysis, Gemini can now mix your drive information with public information to supply insights, summaries, and pattern evaluation. I discovered this replace round gems, significantly noteworthy.
[01:04:59] Paul, I do know we have been [01:05:00] relying increasingly more on gems in a few of our work, and it looks like having these embedded in workspace apps could be fairly helpful.
[01:05:07] Paul Roetzer: Yeah, it may very well be monumental. I, I, once more, I may assume the persons are just a little bit extra acquainted with GPTs, customized GPTs with ChatGPT. ‘trigger yeah, they simply have an even bigger consumer base.
[01:05:17] However I discussed Ada, the educating assistant I created to assist with the event of my programs. And I educated it to ask questions, problem my considering, strengthen concepts, brainstorm ideas, conduct analysis, advocate actions, after which to all the time be fixing for the shopper. So I educated it on a reasonably important system immediate.
[01:05:34] however what I did is I constructed a customized GPT model and a gem from, from Google, on the very same system, immediate and very same information base. And the primary few programs I used to be creating, I would really like speak with each of them. I used to be form of saying, okay, here is my define. What do you assume? How would you enhance it?
[01:05:52] And so I used to be going forwards and backwards and by in regards to the third course, I simply stopped utilizing the customized GPT. The gem was, yeah, far superior to the [01:06:00] customized GPT, at the very least on this occasion. and so. Yeah, I would not sleep on gems each as like simply an experimental factor, like go mess around with them if you have not constructed one, as a result of the Gemini 2.5 professional mannequin they’re constructed on is unimaginable.
[01:06:15] and their potential software within the workspace now’s, requires big change administration and training and coaching and like integration into it. And I feel that is the place most enterprises have this like, big alternative forward is like, simply resolve the plain issues higher than all people else. Like you do not have to determine every part out about AI and resolve tremendous clever, all this stuff.
[01:06:37] Simply take pocket book, LM and deep analysis and customized GPTs or gems and like educate your crew immediate properly, like there is a large cane available in each single firm crew division by simply doing these few issues rather well.
[01:06:52] Mike Kaput: Yeah, for positive. And we have been working with, , at the very least one firm on serving to them out with some GPTs and even only a very [01:07:00] primary pilot undertaking, not solely created big transformative wins for them with nothing that I might take into account rocket science, extra simply training coaching and ongoing utilization and monitoring.
[01:07:11] And in addition that grew to become an enormous win internally that has now impressed different groups and executives to maneuver ahead additional with ai. So do not, yeah, do not underrate not solely how a small undertaking can do large wins, but in addition what it may well encourage different folks
[01:07:26] Paul Roetzer: to do. Yeah. Do not, do not overcomplicate this after which share the wins.
[01:07:29] That is one thing we have gotten significantly better about internally at Smarter X, is like, we now have a AI playground the place folks submit like, cool issues they’re doing every single day. and it does it identical to, oh man, that is an excellent concept. Like, I am gonna use that in buyer success, or I am gonna take that and apply that to advertising.
[01:07:44] It is it, . It would not, you do not, once more, do not overcomplicate it. Simply do the plain issues properly and have a change administration plan in place that like trickles that right down to everybody within the crew division group.
[01:07:57] Apple Would possibly Use Anthropic or OpenAI for Siri
[01:07:57] Mike Kaput: Alright, subsequent up, apple is contemplating a [01:08:00] main shift in technique. They’re contemplating outsourcing collection core AI to Anthropic Cloud or openAI’s ChatGPT.
[01:08:07] this could make a dramatic reversal of its longstanding dedication to constructing its personal language fashions. In line with Bloomberg, apple has requested each corporations to coach variations of their fashions that would run on Apple’s personal cloud infrastructure as a check if it strikes ahead, that will mainly be an open admission that Apple’s homegrown basis fashions have not saved tempo with rivals and that Siri’s lagging efficiency wants a close to time period repair.
[01:08:33] The potential swap comes after a messy shakeup. Inside Apple’s ai org. Management of Siri has been handed from ai Chief John Gandia. To the crew behind Imaginative and prescient Professional and inside exams reportedly discovered Anthropics. Claude outperformed Apple’s fashions resulting in severe talks about licensing it. So Paul, nothing has been determined right here, however this simply appeared like extra bother for Apple.
[01:08:57] They only misplaced a key engineer engaged on [01:09:00] these basis fashions to meta for about $200 million. I additionally discovered it attention-grabbing in one of many studies from Bloomberg, they stated that quote, apple is understood in lots of instances to pay its engineers half and even lower than what they will get on the open market. So I do not know what is going on on right here, however this does not appear to be it bodes properly.
[01:09:19] Paul Roetzer: Yeah, I imply, if there was one other innovator’s dilemma, you possibly can do a chapter on Apple and ai. yeah, I do not understand how they compete truthfully. Like, I feel that, I do not even know that Acqua hiring and even straight up acquisition works be as a result of it is so counter-cultural to, to Apple. They don’t seem to be gonna pay $300 million for prime AI researchers.
[01:09:41] Like, it will simply basically change the best way Apple does issues. And it’s not an organization that pivots their tradition. So I do not know, like I feel that licensing in all probability ultimately is the play for them. I, I, once more, like you possibly can go acquihire minstrel or, or, or a straight up purchase [01:10:00] perplexity or perhaps even like go after Anthropic, which, , appears in all probability probably the most aligned from a worth perspective.
[01:10:07] Paul Roetzer: You may try this, however why would these researchers and engineers keep at Apple? Like, I do not know. Like, I I, they’re simply not constructed to be like a wartime aggressive AI lab. And so I used to be on the acquisition prepare. I imply, truthfully, up till I began speaking proper now, like, as I am like considering out loud, it’s not gonna work.
[01:10:29] Like they, they would not be capable to preserve folks there. So. I really like Apple. Like I am, every part we do is Apple. I personally, each system is Apple. I have been an Apple fanboy for, , 20 years. I really like the corporate. I do not assume they will compete as an AI lab. So I do not know. I simply, I simply need Siri to work.
[01:10:52] Like I simply, I would like, I do not care whose tech they construct it on, simply make the factor work, so. Proper.
[01:10:59] AI Product and Funding Updates
[01:10:59] Mike Kaput: Alright, [01:11:00] for our ultimate subject, we will run by way of some ultimate AI product and funding updates. Paul, when you have something to chime in with right here, be happy. In any other case I will simply kinda run by way of these as we wrap up.
[01:11:09] So first step, Grammarly is buying Superhuman, which is an AI powered e-mail startup. That is a part of a broader push to turn out to be an AI powered productiveness platform. So superhuman, which was final valued at 825 million, constructed its fame on velocity and design, and helps customers churn by way of emails at Lightning tempo.
[01:11:29] The corporate has discovered itself squeezed as Google and Microsoft have layered AI into their very own e-mail instruments. Their annual income now sits round 35 million. Grammarly, in the meantime is flush with a billion {dollars} in funding from Normal Catalyst, and its rebranding itself past grammar correction and transferring deeper into productiveness.
[01:11:48] So their imaginative and prescient is to make use of Grammarly AI brokers, however superhuman inbox to construct a better communication hub that understands your e-mail schedule and workflows. Subsequent Google Pocket book, [01:12:00] LM obtained just a little little bit of an improve. The AI powered analysis Instrument now function that now contains featured notebooks, that are curated collections of top quality content material from trusted sources like The Economist and The Atlantic.
[01:12:12] And every pocket book combines authentic supply materials with all of Pocket book lms nice options like the power to ask questions, hint citations, create Visible 9 maps, and take heed to AI generated audio overviews. And final however not least, X OpenAI, CTO. Mira Tis Startup Considering Machines Lab has confirmed it, has raised an enormous $2 billion funding spherical led by Andreessen Horowitz and joined by corporations like Nvidia.
[01:12:38] The Startup’s mission stated Mirati in a submit on X is to construct collaborative normal intelligence. Their first product they are saying is coming within the subsequent few months and can embody open supply elements designed to assist researchers and startups constructing customized fashions. Paul, that may be a wrap on a busy, busy week, two weeks in ai.[01:13:00]
[01:13:00] Paul Roetzer: Yeah, and like we stated, there’s actually dozens of issues we did not get to. so take a look at the, this week in AI publication. we, we, we’re again now. I do not, I do not assume we now have any disruption to the weeklies transferring ahead that I do know of. So, as of the second we’re, we, we again on our common Tuesday schedule and I nonetheless have two extra course collection to finalize.
[01:13:22] So. I am form of nonetheless, , locked within the, within the lab constructing programs for the upcoming launch. And keep tuned for the AI Academy by Good Rx, relaunch information. It it will be, popping out, we’ll in all probability speak in regards to the podcast subsequent week. however if you’re an AI Academy member or eager about being one, co coming very quickly.
[01:13:44] Alright, thanks Mike. Thanks Boff. Thanks for listening to the Synthetic Intelligence Present. Go to smarter x.ai to proceed in your AI studying journey and be part of greater than 100,000 professionals and enterprise leaders who’ve subscribed to our weekly newsletters. [01:14:00] Downloaded AI blueprints, attended digital and in-person occasions, taken on-line AI programs and earned skilled certificates from our AI Academy, and engaged within the Advertising and marketing AI Institute Slack group.
[01:14:11] Till subsequent time, keep curious and discover ai.