273 | AI’s civil-war: Anthropic vs OpenAI vs Dept-of-War showdown, Anthropic revenue doubles since January 1st, GPT-5.4 & Gemini 3.1 released, Perplexity Computer launches, and more pivotal AI news for the week ending on March 6, 2026

What happens when AI companies, the U.S. government, and national security collide in public… on social media?

In this Weekend News episode, Isar Meitis breaks down one of the most dramatic developments in the AI industry so far: the escalating conflict between Anthropic, OpenAI, and the U.S. Department of War. What began as a disagreement over AI guardrails quickly turned into a public showdown involving government ultimatums, billion-dollar contracts, and a massive shift in public opinion across the AI ecosystem.

The situation reveals a deeper question that every business leader should be thinking about: who ultimately controls powerful AI systems — the companies that build them, or the governments that rely on them?

In this episode, Isar walks through the full timeline of events, analyzes the implications for the future of AI and geopolitics, and shares what this could mean for businesses navigating the rapidly evolving AI landscape.

In this session, you'll discover:

• What triggered the standoff between Anthropic and the U.S. Department of War
• Why AI guardrails around surveillance and autonomous weapons became the breaking point
• How OpenAI stepped in with a Pentagon deal — and why it sparked backlash
• The impact on public trust, app downloads, and market perception
• What this conflict means for the AI race between the U.S. and China
• Why the government labeling Anthropic a “supply chain risk” could reshape the tech ecosystem
• The broader implications for companies doing business with governments
• Key AI model releases including GPT-5.4, Gemini 3.1 Flashlight, and Perplexity Computer
• Why AI agents, computer-use models, and coding automation are accelerating
• How new research shows AI study tools improving learning outcomes for students

About Leveraging AI

• The Ultimate AI Course for Business People: https://multiplai.ai/ai-course/
• YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
• Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/ 
• Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events

If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!

Transcript

Speaker 0:00

Hello and welcome to a Weekend News episode of the Leveraging AI Podcast, the podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Metis, your host, and we have a lot of interesting things to talk about. The biggest one will be continuing a story that we started last week with the relationship between open ai, anthropic Trump, the Department of War, and many other participants. While I did report on this last week, I recorded the episode on Friday afternoon. As the things were evolving, I didn't have the full context of what was going on, and there's a lot of that happened since. That has significant, I think, implications on how AI is evolving in the world and the impact it has on everything. So we're gonna dive into that and gonna spend a lot in analyzing that situation. We have several model releases, including two models from OpenAI and we have a lot of rapid fire items to cover as well. So let's get started. So for those of you who missed the news about Anthropic and open AI and the Department of War, the quick recap is that Anthropic refuse to change the terms of their agreement with the government to allow the government to use it for what the government calls any lawful use Anthropic was insisting that their technology's not gonna be used for surveillance against Americans and not for completely autonomous weapons. Secretary Hegseth, who is in charge of Department of War warned them that if they do not comply, he will ban them or basically define them as a supply chain risk, which is something that is designated for adversaries who risk national security. And both him and President Trump decided that because Anthropic would not comply, they would be designated as such. Very shortly after they signed a very similar deal with OpenAI and with Xai. And as a result of that, there's been a massive shift in public opinion against OpenAI and towards Anthropic. That's on a nutshell what happened, but I do wanna deep dive into a lot of the components and the specific quotes and what was said and what will happen and where are we now and what are the implications I think this has on a much broader scale and what you need to be thinking about on this topic. So first of all, let's go back to July of 2025. The Pentagon then awarded a $200 million contract, to Anthropic and Google and OpenAI to develop agentic AI solutions for defense use. Now, anthropic was the first one out of them to actually get an approval to deploy in a classified military environment versus a non-classified environment. And they were the only. Company that was allowed to run under the classified environment together with a partnership with Palantir Technologies. Now in December of 2025, the Department of Defense that was rebranded, the Department of War, struck a deal with Google to build a Gemini powered internal LLM that will be called Gen AI Mill. And at the same time, defense secretary Pete Hegseth was pushing for military ai and I'm quoting without ideological constraints that limit lawful military applications. And he declared publicly, and I'm quoting again, the Future of American Warfare is here, and it spells a I. Now on January 9th, HX F issues a memorandum that outlines the Pentagon AI strategy calling for us to become, and I'm quoting again, AI first fighting force. The memo signaled that the Pentagon would only contract with AI companies willing to allow the government to use it for, and I'm quoting again any lawful use, and it called for all the AI suppliers to completely strip any safety guardrails against use in mass surveillance and autonomous weapons, which was under their terms of their agreement with Anthropic, which that's what they requested when they agreed to be a part of this contract. Now he gave a deadline of the end of February for contractors to comply with this request. Now on January 3rd, US Special Forces conducted Operation Resolve where the, it was raiding Caracas to capture the Venezuelan President Maduro and reports later on by the Wall Street Journal. And Axxis revealed that Anthropic Claude model was used in the preparation and the execution of this operation. So Anthropic is already built into and weaved into the most critical operations of US military, both on the preparation side as well, on the execution side of such operations. Now in mid-February, Anthropic was holding its ground basically saying it is not going to cave and they will not allow their models to be used for surveillance against American citizens and or using it for completely autonomous weapons. And a senior administration official told Axios, and I'm quoting any company that would jeopardize the operational success of our war fighters in the field is one we need to reevaluate our partnership with going forward. At that time, the Pentagon already started accelerating talks with open AI to fill the potential gap, if Anthropic will not comply with the new demands from the secretary. Now, at that time separately, hundreds of employees from Google and OpenAI signed an open letter urging the leaders to stand in solidarity with anthropic and refuse the Pentagon's demands. Now on February 23rd, secretary Hegseth hosted Dario Amee, the CEO of Anthropic at the Pentagon in what later people said was a very tense meeting. Hegseth presented Amee with an ultimatum basically to lift the guardrails. Allow the government to use it again for all lawful uses by February 27th, or be designated a supply chain risk. For those of you who don't know what that means, it is something that the US government can use in order to define an adversary company that it has done before for companies like Hui. So it's a Chinese chip manufacturer, preventing any government agency or any supplier to the government from using those chips because it presents a risk to national security. This is very clearly not the case with Anthropic because they have already been used in a classified environment in theater while conducting a military operation. So obviously they're not a risk to the military, and yet that was the threat that Secretary Hegseth has used in this particular conversation. On the 24th of February XA, I signed a formal deal with the Pentagon to use grok in classified systems, having accepted the all lawful purposes standards that Anthropic has refused to sign. Now defense officials at the same time acknowledge that it will be very hard to pull out Claude from the classified systems because it's already very deeply connected in multiple places, and they said it's going to be a very difficult process. Now, on Thursday, February 26th, so a day before the ultimatum expires. And after the Pentagon delivered what described as the final offer, anthropic, CEO, Dario Amede publicly stated the company could not, and I'm quoting in good conscience, accept the terms. Now, they also gave the reasons and the reasons were that the current AI models are not reliable enough for fully autonomous weapons. And that mass domestic surveillance violates fundamental rights given to us people by the Constitution. Now at the same evening, open AI's, CEO Sam Altman sent an internal memo to staff saying OpenAI shared the same red lines as Anthropic and was beginning its own negotiation with the Pentagon. For a deal with them as well, with the exclusions for domestic surveillance and autonomous weapons without human approval. So in the internal memo that Sam Altman sent to his employees, he shared that they're going to align with anthropic on these red lines. In the afternoon of the 27th, both Donald Trump and Secretary Hegseth issued their own statements relating to the situation and relating to the fact that Anthropic would not cave in to the demands from the government. Trump wrote on truth social, the left-wing nut jobs at Anthropic have made a disastrous mistake trying to strong arm the Department of War and force them to obey their terms of service instead of our constitution. He continued and he ended with, we don't need it. We don't want it, and we will not do business with them again. On the 27th when the ultimatum expired, both the president and secretary Hegseth issued different statements about the current situation. President Trump in untruth social wrote the following, the left wing nut. Jobs at Anthropic have made a disastrous mistake trying to strong arm the Department of War and force them to obey their terms of service instead of our constitution. Therefore, I'm directing every federal agency in the United States government to immediately seize all use of Anthropic technology. We don't want it, and we will not do business with them again. He then continued to say that Anthropic needs to support the government for a six month phase out phase. And if they somehow fail to support the government, he will use his entire force in order to take action against them shortly after. Secretary Hegseth, issued his own statement saying, and I'm quoting in conjunction with the president's directive for the federal government to seize all use of anthropic technology. I'm directing the Department of War to designate anthropic supply chain risk to national security effective immediately. No contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with anthropic. Now, I don't want to get into politics at all, but I have serious issues with several different aspects of this.. Some of them I will touch in the end, but the first one that I wanna address right now, both these statements that are official statements of the government related to a agreement with a US company instead of being discussed in closed doors as they should have been released on truth social and on X. Like this is not a letter going to anthropic. This is not a phone call. This is a formal announcement of very severe actions that the government is taking that is decided by two specific individuals that are sharing it on social media. So now let's continue just a few hours later. Still on the 27th, late on Friday night, OpenAI, CEO Sam Altman posted on X announcing that OpenAI had reached a deal with the Department Board of War to deploy each models on their classified networks. Basically taking the position that Anthropic was in. And in addition, following Trump's statement, secretary of War, Pete Hegseth has issued his own statement on X saying the following, and I'll read most of it because it is critical so you guys understand the exact details. This week, Anthropic delivered a masterclass in arrogance and betrayal, as well as a textbook case on how not to do business with the United States government or the Pentagon. Our position has never wavered and will never waver. The Department of War must have full unrestricted access to anthropics models for every lawful purpose in defense of the Republic. I'm gonna skip a section and then he continues. The terms of service of anthropics defective altruism will never outweigh the safety, the readiness, or the lives of American troops on the battlefield. Their true objective is unmistakable to seize veto power over the operational decisions of the United States military. That is unacceptable. Now, I'm gonna pause here for a second and say that Anthropic has given full access to its models, to the US government that has been used in military actions, including in Iran. We're gonna get to that in a second because they believe that this is the right thing to do, and because they believe that the US military can benefit from that, the only thing they have sanctioned is the use of these models against US citizens and completely autonomous weapons because they think the models are not there yet from a capability. So they definitely did not request any veto power on operational decisions. And then he continues towards the end. In conjunction with the president's directive for the federal government to seize all use of anthropic technology, I'm directing the Department of War to designate anthropic supply chain risk to national security effective immediately. No contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic. Anthropic will continue to provide Department of War its services for a period of no more than six months to allow for a seamless transition to a better and more patriotic service, Amer, and then he ends with American war fighters will never be held hostage by the ideological whims of big tech. This decision is final. So before I continue with what happened next, I wanna say one thing, and I don't wanna dive into politics at all, and I'm trying to stay as objective as possible here in describing to you what actually happened. The biggest issue that I want to address in this immediate section is that this is a legal contractual arrangement between a US based company that is worldly renowned and successful, and the US government. This kind of agreement or outcome might be acceptable, but it should not take place on truth, social and X. It should take place in a formal letter, in a closed room, in a phone call, something like this, but not on social media. I think this is a horrible way to conduct any kind of business relationship between any organizations, whether military, government or civilian. But now let's continue. So just hours after these two announcements, OpenAI, Seman posted on X announcing that OpenAI has reached a deal with the Department of War to deploy its models on their classified network, basically taking the place of Anthropic in that role. Now the news that OpenAI deal announced just hours after Anthropic was banned, generated an immediate insignificant public backlash against OpenAI sensor. Tower data showed that Chacha PITI app uninstalled spiked 295% on Saturday immediately after the announcements on Friday night. Meanwhile, downloads to the Anthropic Claude app Rose 37% on Friday, and 51% on Saturday. Ended up with Claude being the number one downloaded free app in the United States overtaking OpenAI, while OpenAI is being uninstalled significantly faster than it did before. Also protestors surrounded OpenAI San Francisco headquarters and created a lot of chalk messages on sidewalks, stating on what they think about the situation, which led Sam Altman to hold an extended ask me anything, session on x. Admitting that the deal has been definitely rushed and that the optics don't look good. And then in the beginning of the following week, March 2nd and third, Monday and Tuesday, Sam acknowledged the following. We were genuinely trying to deescalate things to avoid a much worse outcome, but I think it just looked opportunistic and sloppy. No shit. On Monday night, Altman unveiled a reworked agreement with the Pentagon, adding explicit language, stating the AI systems shall not be intentionally used for domestic surveillance of US persons and nationals. That it's not going to be used for autonomous weapons either. Now while all of this is happening, I want to remind you that the US military forces are conducting military strikes on Iran, and it has been shared that anthropics Claude has been a key tool for the US central command. To successfully deploy such strikes, both in the planning phase as well as the execution phase of this military action. So again, while all of this is happening while the US President and the US Secretary of Defense, or Secretary of War, or whatever his title is right now, are banning the use of anthropic, the actual US military is using it actively at war in order to gain a benefit in that arena. Shortly after the initial OpenAI suggesting that they're going to change the terms of their agreement. OpenAI has also temporarily paused the deployment of its AI technology to the NSA and the Department of War, according to an expo by OpenAI researcher Nolan Brown. This decision to pause the deployment was done to make time for the Democratic process to address the potential surveillance and loopholes identified in the original agreement language, particularly concerning how AI could enable legal surveillance. Now on March 4th, the information has reported that an an internal memo, Dario Amide has basically exploded on both open AI's actions as well as the government. And he wrote a very long and detailed memo, and I'm gonna put the link in the show notes for you to be able to read it. And I'm gonna read a few sections out of that because I think it's important for you to understand at least ammo day's opinion on the current situation. So he starts with, I want to be very clear on the messaging that is coming from OpenAI and the mendacious nature of it. This is an example of who they really are, and I want to make sure everything sees it as is. And then he says they don't really have access to the final agreement, but then he says, we do know the following, and I'm quoting again, Sam's description and the DOW and the Department of War description, gives a strong impression that how their contract works is that the model is made available without any legal restrictions and in quotation marks, all lawful use. But that there is a safety layer, which I think amounts to model refusals that prevents the model from completing certain tasks or engaging in certain applications. Safety layer could mean something that partners such as Palantir, which is Anthropics business partner for this, kind of work tried to offer to us during these negotiations, which is that they on their end offered us some kind of a classifier or machine learning system or software layer that claims to allow some applications and not others. There's also some suggestion of open AI employees looking over the usage of the model to prevent bad applications. Our general sense is that these kinds of approaches, while they don't have zero efficiency, are in the context of military applications. Maybe 20 real and 80% safety theater. The basic issue is that whether the model is conducting applications like mass surveillance or fully autonomous weapon depends substantially on their wider context. A model doesn't know if there is a human in the loop. In the broad situation, it is in for autonomous weapons and it doesn't know the provenance of the data it is analyzing. So doesn't know if this is a US domestic data versus foreign doesn't know if it's enterprise data given by customers with consent or data bought in sketchier ways, et cetera. We also know those in safeguards know painfully well that refusals aren't reliable and jailbreaks are common often as easy as just misinforming the model about the data. It is analyzing an important discussion here that makes it much harder for the safeguards problem is that while it is reliably easy, for example, to determine if a model is being used to conduct cyber attacks from inputs and outputs, it is very hard to determine the nature and the context of the cyber attacks, I'm skipping a section and then he saying that refusals aren't reliable and jailbreaks are common, often as easy as misinforming. The model about the data it is analyzing. So that is addressing one aspect of the safety layer. He goes back to a second aspect of the safety layer. This kind of safety layer stuff that Palantir offered us and presumably offered O Open AI is even worse. Our sense was that it was almost entirely a safety theater, and that Palantir assumes that our problem was you have some unhappy employees. You need to offer them something that placates them or makes what is happening invisible to them, and that's the service we provide. Then he addresses the third component of potential safety cover, which is finally the idea of having anthropic or open AI employees monitor The deployments is something that came up in discussions with Anthropic a few months ago when we were expanding our classified A UP, which is acceptance use policy of our own accord. We were very clear that this is possible in only small fraction of cases, so I'll skip the rest. Basically, that doesn't work as well. Then he is talking about some specific language that OpenAI presumably may be trying to put into their account. But he's saying we actually attempted to include some of the same safeguards as OpenAI in our contract in addition to the acceptable use policy, which we consider the more important thing, and the Department of War rejected them with us. We have evidence of this in the chain of emails, blah, blah, blah. So basically the Department of War is allowing OpenAI to get terms that were rejected when anthropic requested similar terms. He's saying We have evidence in the email chain, off the contract negotiations, and I'm writing this with a lot to do, but I might get someone to follow up with the actual language. Thus, it is false that OpenAI terms were offered to us and we rejected them. Then he is talking about the fact that because all they're granting the government is lawful action and because spying against US citizens is illegal, then there shouldn't be a problem. But then he's adding the following, and I'm quoting again. It is however, completely false as we explained in our statement yesterday. The Department of War does have domestic surveillance. Authorities that are not of great concern in a pre AI world, but take on a different meaning in a post AI world. And then he adds for specific language. Notably, near the end of the negotiation, the Department of War offered to accept our current terms. If we deleted a specific phrase about analysis of bulk acquired data, which was the single line in the contract that exactly matched this scenario we were most worried about, we found that very suspicious. Then he is adding as far as autonomous weapons, that while it is currently the law that humans have to be in the loop for firing a weapon. It's a Biden administration law. This can be changed at Will by Pete Hegseth, and hence it's not a real constraint. Then he goes to Sam Altman's approach and he's basically saying that Sam Altman in his statements is trying to make it look as if he has. But in reality, he's working against them, behind the scenes, conspiring with the Department of War to take their place while trashing anthropic, while trying to make it look like he's the good guy and trying to deescalate the situation, which is the exact word that Sam Altman has actually used. But then he goes further and he says the following, the real reason Department of War and Trump admin do not like us, is that we haven't donated to Trump. While OpenAI slash Greg Rockman have donated a lot, we haven't given dictator style praise to Trump. While Sam has, we have supported AI regulation, which is against their agenda. We've told the truth about a number of AI policy issues like job displacement, and we actually held our red lines with integrity rather than colluding with them to produce safety theater for the benefit of employees, which I absolutely swear to you is what literally everyone at the Department of War, Palantir and our political consultants, et cetera, assumed was the problem we were trying to solve. Sam is now with the help of the Department of War trying to spin this as if we were unreasonable. We didn't engage in a good way, we were less flexible, et cetera. I Want people to recognize this as the gaslighting it is now this was actually an internal memo, and yes, it was leaked. And yes, Dario needs to know that something like this will get leaked, especially in the current way of things. But that was released by Dario Amide, which obviously did not make him a good friend of the government more than he already is. However, while all this is happening anthropic is supposed to go public this year. Anthropic has some very serious investors like Amazon and Nvidia and Iconic, and some really big firms with really deep pockets and really serious connections in the government. Additionally, the information technology industry console and industry group whose members include Amazon, Nvidia, apple OpenAI sent a formal letter to the defense secretary, Pete sef, expressing deep concern over the pentagon's use of the supply chain designation risk in response to a standard procurement dispute. So apparently right now Anthropic is back in the negotiation table with the government. Now can this be resolved somehow after the previous statements that what said? I'm not sure. So let's talk a little bit about the aftermath of what's happening right now, and then I'm gonna tell you what I think about the entire situation. So while all of this is happening, as I mentioned, there's been a big backlash against OpenAI. Millions of people basically started following different hashtags against OpenAI as well as a lot of people has started moving to Anthropic. Now, Anthropic was in a positive swing before that. I told you that I've spent 95% of my time right now on the Anthropic platform and 5% divided by the rest of the platforms combined, including open AI as one of them, and I'm not the only one. So that has been the sense around for a while now. Well, we just learned that Anthropic has hit a 19 billion a RR rate by the end of February. This is up 36% in the past two weeks. We're talking about billions of dollars in revenue. That gets added in a couple of weeks of time. Shortly after that announcement, OpenAI released their numbers saying they're currently outta the 25 billion a RR rate, but it's very obvious that the gap is closing and it is closing very, very fast. went to 19 billion from about half that at the end of 2025. While OpenAI have seen a 17% growth since the end of last year, so at this rate, anthropic will pass them very, very quickly in revenue despite the fact they raised significantly less money and has significantly less resources and employees. So now I wanna summarize my thoughts on the issue, and then we'll dive into lots of interesting rapid fire, like new models and so on. There are several distinct aspects of this that require analysis and requires attention. And I'm going to describe the issues. I'm gonna tell you what I think about them, and then I want you to think about it on your own and what you think about each and every one of these aspects. So number one is, can a supplier to the government and specifically to the Department of War decide how the government can use their products or services? I believe the answer is very simple. The answer is no. The US government can define its needs, and a contractor can decide whether he wants to work under these circumstances or not. Based on my understanding of the situation, the Anthropic original agreement had the specific terms and terms and limitations on terms of use, and this is what Anthropic signed up for. Then the government decided to change these terms afterwards, and it is perfectly fine. Any agreement could be renegotiated, whether between the government or even between two business entities. But if one side decides to change the term, the other sides can decide whether they want to agree to the new terms or not. I'm certain that similar situations happened numerous times before with the government and ended up quietly by the company walking away from the government contract, which is perfectly fine. Meaning if one side decides to change the terms, the other side can decide not to continue under the new terms. This sounds and makes perfect sense to me, and this is, again, probably happened numerous times. It does mean you lose your government contract, obviously, but it should not mean and never did by the way, in history of the United States, that the government should go nuclear and try to destroy the company because it does not agree to the terms. So this is number one. Number two. The opposite side of that story is that the United States is in a race against AI supremacy and global supremacy with China. The Chinese government has full access by definition because of the way that country works to anything Chinese companies develop without any limitation to how they're going to deploy it, they will definitely deploy it to spy against Chinese citizens, and they will definitely deploy it in any military way that they think that will benefit them. Meanwhile, if we limit the US government's access to the top US technology, we are limited the US ability to stay ahead in the race with China. Whether it's an economical outcome or a war kind of outcome. In both cases, the country that will have full access to AI has a better chance of winning. So by limiting what the government can do with the best AI capabilities, these companies are potentially limiting the US ability to win the future war again or economic conflict with China. Now, aspect number three is what does the designation of Anthropic as a supply chain risk actually mean to the US technology sector? Currently on paper, any company that works with the government cannot work with Anthropic, period. This is what hegseth and the President said this includes companies like Microsoft, Google, Amazon, Oracle, that all are supplying services to the government and many other companies as well. But what that means on paper is that these companies will need to choose between keeping their government contracts, which are worth billions and billions of dollars, or keeping Anthropic alive because if Anthropic cannot use AWS Google Cloud or Microsoft Azure, they have no means to deliver their services and they're basically done. In other words, if this moves forward as it's defined right now, it is a death warrant for Anthropic. Now we are talking about the company that is maybe the number one AI company in the world today. If not, it's definitely one of the top three or four, which means it is an asset for the United States, and if the US loses Anthropic, it will be a very important asset that they lose in the race against China. Now is also very clear that the current usage of Anthropic by the Department of War has been extremely valuable and has been used actively in the two military engagements of the US military in the past few months, which means there are not a supply chain risk, but rather a US national security asset. An important asset because it's been used again effectively in two out of the two military engagements the US military had in the past couple of months. But the bigger question that is connected to that is what do these actions mean to any company who wants to do business with the United States government? It doesn't matter whether it's foreign or domestic. It means that at any given moment, any given company is at risk, that the government will decide to change the terms of the agreement. And if you do not agree, the US government will take action, any action that it can to destroy your company. I do not think this is a good way to conduct business, or I do not think it's a way that will encourage companies to work with the US government. I also do not think this is how the US government and definitely not the US people want to treat existing and potential business partners of this country. Now the last topic, which goes back to my comment in the beginning about doing this on X and truth social is how politicized the whole situation became this could have been easily resolved in closed rooms, and instead we get President Trump saying things like the Leftwing nut jobs are Anthropic, or Hsf saying a masterclass in arrogance and betrayal. Dario saying we haven't given dictator style praise to Trump while Sam has. All of these are not helpful to get to a simple solution, whether the solution is finding a way to allow Anthropic to continue to work with the government or not, but it could be resolved in a much more normal way of doing business without getting into these kind of statements. In general, I think the decisions that impact national security or the future of any country, including the US, should not be based on political preference. Left wing, right wing, whatever agenda you're in, and definitely not from a personal rage or vengeance kind of point of view, which all of these statements sounds very much like that, and it should not be made by one or two people. This is why we have a system in place. This is why there's parliament, this is why there's a judicial system. This is why this country has thrived for so many years because there are checks and balances and one person cannot do these kind of things. And I'll quote Norm Brown from OpenAI that summarized it in a very nice way. I'm afraid of a slippery slope where we become a customized to circumventing the democratic process for important policy decisions. When there is bipartisan support, and urgency, I have faith that the government can act quickly. So to summarize this whole situation, I really hope that this is gonna get resolved in the right way, meaning finding a way to allow the US government to use the most advanced technology while keeping US citizens safe. And while not deploying fully autonomous weapons, I really hope that nobody, no country in the world, will deploy fully autonomous weapons using AI because that raises a lot of red flags and questions and talking about the slippery slope of going, of allowing autonomous weapons, fully autonomous decision making processes that end up with deploying weapons will end up with nuclear weapons as well. And then we all see in the movies where that can go. Now to our rapid fire items. And we're going to start with the release of several new models. So first of all, OpenAI released GPT 5.4, which is currently on many of the benchmarks, the most capable AI model. Out there, it has advanced reasoning, coding agent capabilities, native computer use functionality, and a lot of goodies baked into one new model. Now, in addition, scoring very, very high on multiple evals, including the GDP valve, which is a benchmark that is spanning 44 different occupations that have pushed it from 70.9 on that eval for G PT two to 87.3 on GPT 5.4, which is very, very significant. It means it's very good across 44 different actual occupations of people. GPT 5.4 is also the first generation model with native computer use capabilities. And it achieves a 75% success on the OS world verified benchmark that benchmark checks how desktop environment is navigating and it is doing it better than humans. That scored 72.4 on the same benchmark. So it's the first model that actually does better on using computer than average humans. It is also 33% less likely to make false individual claims and 80% less likely to contain any errors in full responses compared to 5.2. They also have significantly improved its tool usage efficiency, so it now can use tools significantly more effectively than previous models. It is relatively expensive. OT as expensive as the top off the line anthropic model. So GPT 5.4 API, input tokens are $2.50 for million tokens and $15 for million output tokens. This is up from 1 75 and 14 on GPT 5.2, but it's not a crazy jump up. Now in the Artificial Analysis Intelligence Index, which is a company that independently checks all the models across a very wide range of topics, it currently ranks number two out of 116 with a score of 57 ahead of all the models that you know. The only model that's ahead of it is a very unique, specific version of Chachi PT that is not actually available to the public. However, to achieve this score, it used about 120 million output tokens, which is about 10 x the average for the other models that has been evaluated through the same evaluation. This means that while it's not crazy expensive for a million tokens, because it is reasoning so deeply about everything it's doing, it is actually consuming a lot more tokens than other models. So you will pay a lot more to get the better answers using this model. And this model also has a much bigger context window. So it has a 1.1 million tokens context window, which is up from 400,000 in GPT 5.2. So a very big jump more than doubles the context window, which is a very, very big deal. Now. In addition, OpenAI released another model, which they call GPT five three Instant. And the goal of that is to improve the conversational quality by reducing unnecessary refusals cutting defensive statements, and delivering more direct contextually relevant answers. It also supposed to reduce hallucinations by 26% when using tool access, by using web access, and 19.7% when relying just in on its internal memory. And the main goal of this model, and I assume 5.4 as well, but the focus on this model was clear was to make it more humanly conversational. There has been a very serious backlash on how robotic un warm and unwelcoming is the conversation with current GPT models compared to, let's say, GPT-4 oh and definitely compared to anthropic. And this is a big aspect that OpenAI is trying to fix. And the reason I think this is very important is because when you think about it, all these models are now very, very capable. The differentiator between them becomes more and more your personal preference versus the actual capabilities of the model itself. And if you like using one model better than the other, just because it feels better to work with, there's a much higher chance you're gonna use that model, even if the other model scores two percentage points higher on one evaluation or another. Another big announcement this week was Google just launched Gemini 3.1 flashlight. It is currently the most cost effective Gemini three series model, and the goal is to allow users to enjoy really powerful AI models while working significantly faster and cheaper. So it's currently priced at 25 cents for million input tokens and one and a half dollars of output tokens. Again, that's exactly one 10th of the new GPT 5.4. Now, while it is obviously not as powerful as GPT 5.4, it is the number one model at the tier and class and level it is. And it is very good at reasoning across multimodal inputs so it knows how to see, read and understand both images, video, audio, and text. such as all the other Gemini three class models. Now the goal of this model is connects directly to what I said before as far as the amount of tokens that the AI is going to use. And the idea here is that we need to stop thinking about the cost per token, which is almost meaningless, and we need to think about the value per task. Basically, how much am I willing to pay to get this task done successfully? Because the price per token doesn't actually tell you how much you're going to pay, because that depends on the amount of tokens that the model is going to use. So this model is supposed to optimize exactly for that. A good quote about this came from the head of AI at Latitude. Colby Nottingham, who said Google's models has demonstrated unparalleled instruction following capabilities and speed in its class achieving 20% higher success rate and a 60% faster inference times than other previous models. It's enabling latitude to deliver sophisticated storytelling to a much wider audience than would have otherwise been possible. So again, this is the direction that we've seen a lot of the models going, and this is just a very good example of that, of taking a bigger model, distilling it down, keeping most of its capabilities and making it significantly faster and cheaper. And I'm sure we see a lot more of that. But maybe the most interesting announcement of this past week, or actually week and a half in this particular case is that Perplexity Launch Perplexity computer, which is an advanced agent agent tool that is only currently available to the people who are paying $200 a month for Perplexity Max subscription. But it can autonomously execute complex workflows on your computer. Basically anything you can do. Now, the biggest difference between this and the other computer use tools that we've seen from Claude and OpenAI, et cetera, is that it has access to 19 different AI models and it can switch between them to get the higher efficiency and the best results, and it can try multiple of them at the same time, while giving you the best output. So instead of you having to pick which tool you want to use, it can help you do that and try different things until it gets the work done. Now per perplexity, this solution can execute detailed tasks for hours and even weeks based on their statement. Integrating with business tools, monitoring email calendars, running schedule jobs, and defining proactive triggers for different things to allow the system to run effectively. This is part of perplexity push to switch from a niche consumer product to a more enterprise product, apparently the company right now has tens of millions of active users and about $200 million in annualized revenue, which is not at the level that we hear from the other labs. But for a company that exists for less than three years, that is extremely impressive. Another interesting release this week is two new skills from Anthropic. One is called simplify skill, which is just forward slash simplify used in cloud code that is reviewing modified files and looks for related code reuse, quality and efficiency. And it basically does three parallel review agents to aggregate findings and apply necessary fixes. So it's a tool that helps you check what already happened and make fixes to it. And the other one is forward slash batch, which is a skill to orchestrate large scale called changes in parallel. So more push from potentially the most capable AI coding platform in the world right now, which is Claude Code. And another cool feature that they just announced this week is voice mode for cloud code, which basically you can type forward slash voice and start chatting with cloud code instead of typing. It is currently rolled out about 5% of users, so if you don't have it, don't worry. It will probably roll out over time. I don't have access to it at this moment yet. I really hope to get it because it'll be interesting. But that being said, the initial feedback from users who do have it is that it is not as good as talking to, let's say, Chachi pt, that it's less conversational and less accurate in understanding exactly what you're saying. Now, I have been using both whisper flow the built in Mac voice typing capability inside of everything that I do, including Claude Cowork and Claude Code. So for me, it's not a big deal. I don't see a benefit of it reading it back to me. I actually want to just speak to it and have it do what it needs to do, and then I can quickly skim through what it's actually writing. However, I do see this as a very powerful capability combined with what they announced last week, which is remote control. So, because I now can control cloud code on my computer from my phone, and now adding the voice capability means that I can talk to cloud code that's running on my computer while I'm driving my car or spending or waiting for my son in the parking lot, to come back from his soccer practice. So this is a very powerful capability that it is enhancing our ability to work with Claude Code in new environments. Another launch this week comes from Cursor and they introduced what they called automations, which is a new framework that automatically launches coding agents triggered by code-based changes, slack messages or timers, which basically means that instead of you having to define which agents run and do what, the system orchestrates all of that for you. I can tell you as somebody who's currently running. Three to six projects at any given time. In parallel, it gets relatively intensive and confusing to try to manage and see exactly what each agent is doing at any given time. And having an agent that will allow me to manage all these agents could be very, very helpful. So this is the direction I think the entire world will go. And currently, cursor has developed this for their development environment. So the problem that they're trying to solve is that engineers are now overseeing dozens of coding agents and they're just overwhelmed by the complexity of what's happening. Automations, which is the new feature, breaks. That dynamic by triggering the agents automatically and looping the humans only when necessary. And from releases of features and models to a hardware release and from a company we haven't spoken about in a while, which is Apple. Apple just released M five PRO and M five Max, which is their latest chips that is built specifically to push the capabilities of anything on the computer, but most importantly, AI capabilities. These chips offer four x peak GPU compute over the previous generation of their processors. In addition, M five Max is providing up to 35% faster graphics compared to M four and M four max and 20% higher general graphics performance than M four max. So overall step forward on Apple with the goal for it to be able to run AI locally on different Macs. Now, this is currently aimed specifically for the MacBook Pro and already a lot of people started asking when they can get that processor on Mac Minis to run an army of open claw agents on it. I'm sure that's coming very, very quickly. Another interesting development this week from a technological perspective comes from Black Forest Lab, which is one of the most advanced open source visual generation models, and they have developed a new way to train their visual models, and they're creating this as an infrastructure so other people who want to train visual models can also do this significantly faster. So their new framework that they called Self Flow achieves nearly 50 x reduction in training steps compared to traditional methods that existed today. Meaning if you wanna train a visual model to your company's needs, it will take you 50 x less time with less examples, less steps in order to get to a fully trained model that does exactly what you need to do. This is obviously a big deal because they're not just using it for themselves. They're also allowing other companies to do the same, which is aligned with what we're seeing in now we have seen similar things coming from other aspects, such as text and so on, and code of being able to train faster. 50 X is very, very significant and I'm wondering if these kind of technological breakthroughs are available in other aspects of AI as well. We will see huge leaps forward in the ability to create new, better models than we've seen so far, and it's already been a crazy race and a very, very fast pace of releases of new capabilities. Now many, many companies are building AI based solutions right now that just take the existing models and build on top of them. Two interesting pieces of news about that one, many VC companies are basically saying right now, they're not going to invest in such companies unless they have real significant moats, which most of these businesses don't. So if you're planning to raise money, you need to be well integrated in some real domain expertise beyond just building a product on top of an existing model. But the other interesting thing is that Stripe has just came out with a new solution that enables companies that are using third party models through the API to charge and upgrade on the tokens. So how does this work? If you are charging for a service that uses, let's say, OpenAI or Anthropic or doesn't matter, and you know you're paying $2.50 for every million tokens, you can now charge your customers a markup on that number. So let's say you're paying $2 and 50 cents, you're gonna charge your clients $3 for every token that they consume using your service. And that more or less guarantees you're not gonna lose money when they're big spikes. They're also announced their own AI gateway, meaning through them you can connect to all the leading models similar to what you can do with other existing offerings. Such as Vercel and open router. But the cool thing is, regardless of which one you're using, you can now charge per token while doing an upsell on that without having to develop that functionality yourself. Another cool feature that was launched this week comes from Notebook LM, which is one of my favorite tools. And they just introduced what they called cinematic video overviews. So there were video overviews before in notebook, and it looked like a part, like a narrated PowerPoint presentation. It basically took segments out of the content it had access to in its dataset and created a narrated PowerPoint presentation, which was definitely very cool. But now they've added the capability for notebook lamb to generate the videos on the fly based on the needs in order to deliver the points in the most effective way, meaning it is actually using Gemini as a thinking model of what will be the best way to describe or to portray in a visual perspective what is actually happening in this segment of the summary. It then uses nano banana in order to create images and VO three in order to animate those images, including voice, in order to create the final output. This is currently available only to the top tier users of Notebook Elem. So not to the free tier and not to the basic tier, but just to the people who pay for the max version of Gemini. But I think it's a very cool feature and if it's similar to everything else we've seen from Google, it'll find its way to the free version, or at least to the basic version of Notebook LM as well. Either way, this shows you where the world is going, where the AI on its own can generate significantly sophisticated video presentation of a topic without getting any inputs from the human and still generating very solid results, and it's just gonna keep getting better and better. And then the last topic I wanna cover today has to do with using AI for learning. I have been. Allowing my kids or actually pushing my kids to use the learn mode of Gemini and OpenAI inside their homework. And when they're practicing for tests, OpenAI have just unveiled a learning outcome measurement suite, and they have done this in a collaborative work with several different universities and they have measured over 300 college students using the study mode feature. And they were trying to show whether that actually improves the learning process of these students based on the results they got on the tests. The university from Estonia who participated in developing this tool and the evaluation, tested this on over 20,000 students, ages 16 to 18 for several months. And the outcome is very positive on several subjects and not so impressive on other subjects. So, as an example, in microeconomics, those who've used the AI to help them study has scored an adjusted mean of 63% on 63.7% on the test, while the people who didn't use it scored 55 on the test and on neuroscience, students who use the study mode got 61% while the students who didn't got 56%. So not as big of a spread, but still definitely better results. I am a big believer that AI is the biggest gift that we got to the ability to provide individualized study help to any person, young or old, to learn any topic while following their needs. How fast they want to progress, what is the best inputs to help them move faster. Whether it's watching a video, reading text, taking assessments, playing games, and the AI can spin this up on the fly to help people learn faster. This is the first study I know of that's showing this actually working on just a simple outta the box study mode that exists and available to every one of you and every one of your kids. So my suggestion is use it when you're trying to learn stuff. That is it for this week. This has been a very. Intensive week from a political perspective and also with releases of big new important models. We'll be back on Tuesday with another fantastic how to episode where you can learn how to use AI for specific use case. For now, if you have been enjoying this podcast, I would really appreciate it if you can share it with others and give us a review on Apple Podcast or Spotify, wherever it is that you're listening to this podcast. And I hope you have an amazing rest of your week and I'll see you again on Tuesday.