320 | AI cancer curing breakthrough, context is taking main stage, huge funding rounds, and more important AI news ending week of August 21, 2026

27:44It means that AI will learn more and more tasks and more and more ways to completing those tasks. And we learn what are the best ways to complete those tasks. And we'll get to the point that humans will never be able to do the tasks better than AI. Now, in this particular case, these are four mundane tasks that are repeatable things we do on our computer that probably most humans will not like to do anyway. But that being said, it will get to the point that we'll have to find the things we can actually do. And I'm not sure what that is at this point. The third example that we're going to dive into is from Harvey, which is a legal platform that has been around for a while.

28:19They're extremely successful. They've raised crazy amounts of money. Their revenue is skyrocketing and they've just launched Harvey 2. So the big problem that Harvey 2 is trying to resolve is that every lawyer has to rebuild the context for every time he's working, upload the documents, re-explain the matter, restate the preferences. Harvey too creates spaces that pre-populated with matter context, documents, parties, permissions, ethical walls, all the stuff that it needs, all the stuff the lawyer would need in order to build a case or go through a thing. And it's also looking at the task history, basically Harvey's memory of how an individual lawyer or how a specific firm works.

29:01And it combines these two together to deliver faster, better, more context aware knowledge to the outcome that it's doing. In the product announcement, the Harvey team said, the work carries history and context. Your process carries preferences. Harvey 2 starts with both. Again, even what they're calling process and preference is still context, It is the knowledge to the AI on not the background information, but how to do the actual work, how you do the actual work. So what does this tell us? It tells us, as I mentioned in the beginning, that context is the most important aspect for success in the current AI environment that we are in.

29:41It is not about picking open AI versus Anthropic versus Grok versus Gemini. All these tools are now really good in most tasks. It is about giving the AI the right context in order to understand the task better, in order to understand how you, how your company, how your process, how you individually do the task and learn from that in order to adapt to your working environment and provide you better results, but still aligned with the way you want to work. On one hand, I think this is amazing because it enables companies to accelerate the way they're adopting AI with significantly less friction.

30:17And on the other hand, I think it's missing a much bigger point. Trying to work the way we work today is the easiest from an adoption perspective, right? The AI is not asking you to change the way you work. It is just going to help you in the way you work today. The reality is there might be a significantly more efficient way to do this, which means we will need to give AI more freedom to connect the dots on a much higher level and then help us figure out a better way to do it than the way we're doing it today. I think there are more and more individuals and potentially few companies that are going down that path.

30:52Rethink the way you work. I'm definitely doing this. I'm questioning everything all the time. Does that make my world easy to manage? 100 % not. Does it allow me to do really big breakthroughs every month or two? Yes, 100%. Because I'm not necessarily looking to replicate with AI the things that I'm doing right now. I'm looking at what the goal that I'm trying to achieve, and I'm trying to let AI figure out the best way to do that. But the bottom line is that I want you to take from this segment is that you need to start thinking about context. How do you make the data in your universe accessible to AI in a safe and controlled way?

31:28I'm not saying let every employee in every company connect anything they want to the AI. It's not what I'm saying. That's a catastrophe that is just waiting to happen. What I am saying is you need to find the right ways to give AI access to the relevant pieces of data all the time. So you don't have to figure out what it needs to know when, so it can connect the dots for you and help you make better decisions and drive significantly higher efficiencies from your current resources. And now to rapid fire. As I mentioned in the beginning, we're going to start with some safety and security thing. Greg Brockman expressed serious concern this week that open source models, and he specifically talked about ZZAI's upcoming OpenWay GLM 5.3, which was just released this week, has extremely advanced cyber capability and is likely to, and I'm quoting, significantly accelerate the threat landscape.

32:17What he was talking about is the fact that all the new models, obviously the new Chachupiti Astra that is now being stopped from future development, obviously he is focused on what he knows internally, that is Astra, which is the next model that OpenAI is supposed to release, that they just stopped its training run in order to build better and safer chains to hold it in place and be able to control it. And what he's saying is that while they're doing it, the open source models are not. And based on Greg Brockman and many others, these tools have extreme cyber capabilities that are, and I'm quoting, only a few months behind the frontier.

32:54And it is likely to, and I'm quoting again, significantly accelerate the threat landscape. To explain why this really matters, let's look at Safer AI Cyber Gym Evaluation, what they do is they test AI for different cybersecurity and other risks. When they were trying to test Claude Opus 4.7, the model refused basically every bad thing they were trying to do with it. Literally everything to the point they had to stop the test because it was meaningless because they found no way to allow the model to give them the negative information that they were trying to get. like either that involved cyber as well as the ability to generate different kind of biological or chemical weapons.

33:36On the flip side, GLM 5.2, the previous model from Z.ai, did not refuse any of the offensive cyber attacks that they gave it in the evaluation. So basically every cyber attack request that they give it, it was willing to perform. And it also comes with open weights, meaning it can be better tailored and customized to do these kind of things because it is an open weight model. Now, what does that mean? It means that people with bad intentions will use these models, these open source models to do really bad things. Do I think threat actors will use this? Of course they will, just like any other software they have access to.

34:16Do I think it will be a significant change in the threat landscape? Absolutely not. So he thinks, so what does that It puts us in a world where cybersecurity will become a much, much more complex thing because the AI that will be available to anybody who wants access to it will be there and will allow people to design and perform really significant cybersecurity attacks. And the defenders may not have the same efforts and the same capabilities or at least the same timelines to prepare and block all the issues that they have right now. the issue in the Greg Brockman blog that he wrote, he talks about his own personal blog and how he found in a few minutes, a lot of loopholes, just because there's old technology and things that he connected through the years.

35:02And he's saying it's the same thing with every software or every company out there that has technological debt that makes them exposed that so far was acceptable and becoming unacceptable moving forward. So he's saying every company has just a few months to get its act together and find the loopholes and defend its systems much better. Otherwise, it will be extremely vulnerable to cybersecurity attacks. Staying on the topic of security, but from a completely different angle, OpenAI just launched what they call private safety processing. And the idea is to make sure that companies' data is never exposed to human eyes inside of OpenAI.

35:38So what private safety processing is a new safeguard designed to identify misused patterns across multiple AI interactions without compromising OpenAI's zero data retention commitment for API customers. In my software company, Databreeze, that does invoice vouching at scale using AI, this is a very critical component for us, right? Because we do a lot of AI agents that are running through the process that helps companies go through, analyze, vouch, and reconcile thousands of invoices without human interaction, with humans just approving the step in the end. We obviously use the zero data retention from OpenAI API to make sure that the data from our customers and their customers is not going anywhere.

36:25That being said, in specific more advanced agentic cases, the AI companies want to be able to understand what's happening because it's a much longer horizon, more complex process. As an example, Anthropic has announced that the data that goes into long horizon tasks in Fable 5 is going to be evaluated, which led to almost zero adoption on the enterprise because nobody is willing to take that risk of which data goes where, where it's going to be stored, who's going to be reviewing it, who's going to be seeing it. So this approach by OpenAI, the goal is to still help them understand different patterns in long horizon, big data kind of queries and processes without breaking the promise of zero data retention.

37:07So kudos to OpenAI for this new feature. I assume we'll see the same exact thing from other labs as well. Another, now two new features, capabilities and releases this week that are significant. The first is, I don't know what happened this week, but we have three really significant releases from audio companies all this week. So Pika Labs have dropped four Frontier audio models, soundtrack, music, SFX, and speech. It's priced up to 20 times cheaper than their nearest competitors. It allows to generate long segments of each and every one of these things. So either soundtrack or music or speech and also doing sound effects.

37:47And they're just making really good models at a really cheap entry rate. Another big news from this week when it comes to sound, Whisperflow, which is a company that many, many, many people I know use in order to do voice dictation. I started using them when it started the hype. And then I switched to just internal speech recognition inside of my Mac. I don't think it's as good as Whisperflow. I think Whisperflow is much better, but I think it's good enough for my needs. And so I switched back to a free model that doesn't leave my computer versus a paid model that takes everything that I say to type into every piece of software that I use and send it to a third party.

38:23But anyway, Whisperflow just raised$280 million at a$2 billion valuation. To put things in perspective, nine months ago, their valuation was$700 million, which is$2 billion, meaning it's almost tripled from the beginning of this year to now. They also unveiled Kanto, which is a proprietary speech model, which is engineered specifically for noisy real-world conditions, meaning you can voice type very accurately regardless of the noise around you, which is very helpful if you're like me and you don't type anything anymore and you voice type everything. Another announcement this week came from Cartasia, who shipped Sonic, which is now ranked number one on the Speech Arena leaderboard for both naturalness and speed, with sub-90 milliseconds latency across 40 -plus languages.

39:10And they also added a layer of emotional intelligence in order to make it sound more realistic. And enterprise customers, as part of their release, are reporting 2.9 % conversion lifts and 12.2 % engagement increases just because of the improvement in the models themselves. So big boom on the voice understanding and voice generation side of the world. Staying on the topic of what AI can do these days, Asana, just according to a promotional segment from OpenAI, Asana used OpenAI Codex to remove an outdated testing system called Enzyme in two calendar weeks, a work that was originally estimated to take five years and cost about$6 million.

39:54dollars. The project cost them$12 ,000 running AI models. And this is obviously a very, very significant improvement over the assessment of humans doing the work. One of the things that I talk to leaders of companies right now is that you have to reconsider the ROI of projects or initiatives that you thought, or maybe even didn't think of running because things that were cost prohibitive or the ROI just wasn't there or the risk in the ROI just wasn't there is now there because you can now do things that were impossible to do before or would have cost you crazy amounts of money and hence you didn't even consider them and now you can do that.

40:32This applies from things like this, like removing old system and replacing them with new one, or with potentially completely changing your business model because the way you deliver things can change and it couldn't change before. So this is just food for thought. Another thing that was announced this week comes from Slack. Slack just introduced Slack code. It It is a new feature that brings agentic software development out of the private browser tabs and into dedicated team channels where humans and AI agents collaborate transparently. So how does it work? Basically, in those code channels, Slack code introduces project-based code channels that automatically spins up when a team tags a coding agent, and then the coding agent will actually go through the conversation.

41:14It will go through different code diffs, and it will build whatever it needs to build as part of the conversation with the people in that channel. The idea is taking vibe coding the way we already know it and love it and making it a much more team effort where multiple people can participate in the process. I think this is really smart. And I know that because of the huge distribution Slack has, this is going to be very interesting to see. What I don't know is how the companies will control and manage and deploy and support the code that is generated through this mechanism. them, but I'm sure the infrastructure for that will be developed over time as well.

41:50Now, connecting one more point to the way we started the conversation today with the healthcare side of thing or the research side of thing, Anthropic just published a research this week that has demonstrated breakthrough capabilities in accelerating drug discovery by designing ProteinBinder with 93 % success rate. Basically, 14 out of 15 actually work correctly. They also achieved heat rate of 22 to 35 % compared to industry standards of 10 to 15%. Now, the AI also analyzed complex chemistry data in under 25 minutes, a task that typically takes chemists hours or days to complete. Now, what does this mean?

42:29I'm not a chemist. I don't know the exact understanding of what it means. But what it means is it means that AI can now do really advanced, really complex chemical analysis and chemical design in a way that beats humans by a very big spread, both in means of time and in means of quality of work and success rates. Going from healthcare to another topic where AI is changing the way we engage with the world, a new RTB House survey of over 1 ,800 consumers across US, UK, Japan, and France revealed that shoppers are increasingly trusting AI tools to assist with purchasing decisions. 42 % of US millennials are willing to let AI agents buy items up to if returns are guaranteed within seven days.

43:13However, over a third of all consumers still want human review, basically your approval before AI completes the transactions. That sounds really low to me. Only a third of people will to verify whatever the AI is buying. I'm not sure about that, but I definitely can see how more and more people are trusting AI with helping them find the right output. Now, what makes it really interesting is that AI currently ranks more trustworthy than influencers, newspapers, major media outlets, and social platforms like TikTok and Instagram, with 44 % of consumers trusting AI compared to much lower levels on these other platforms.

43:48The only thing that ranks higher than that is friends and family, which rank at 50. So that's getting close to that as well. Now, another interesting thing that has been released this week that is supposed to dramatically boosts productivity of all the users is replit made a big announcement on august 18 and they introduced what they're calling free mode now the name is really confusing because it is not actually free it is under the 20 a month plan but it lets subscribers create 30 times more code for the 20 that you're using how are they achieving that they're achieving that by integrating chat GPT 5.6 Luna and integrating it into everything they know inside of Replit.

44:305.6 Luna, OpenAI's ways to try to compete with the really low prices of Chinese models. So it is a very capable model at a very cheap price. And by integrating it into Replit, they're now able to perform these tasks again at a 30x higher efficiency than they were able to do before. So huge impacts across the board from shoppers, consumers, science, basically any aspect you can think of. AI is making really big waves even just this week, which leads to crazy, stupid amounts of money that is being poured into anything that has to do with AI infrastructure or any significant thing. So I'm going to throw a few numbers at you.

45:12All of them are from the last two weeks. Most of them or from this past week to tell you how crazy the financial investment is right now when it comes to building AI capabilities. The first one comes from NVIDIA. NVIDIA has partnered with six Wall Street largest financial institutions to establish financing platform and to deliver over 500 billion dollars in third-party capital to develop AI infrastructure. If this doesn't sound like circular financing. I don't know what is, but the idea is that these companies will push money to third-party companies who want to buy NVIDIA compute in order to deliver AI capabilities.

45:55Again, this is half a trillion dollars of investment committed to building AI infrastructure to drive more money to NVIDIA. Staying on the really crazy numbers, Stripe just agreed to acquire OpenRouter for$7 billion. We reported this a few weeks ago, so it's not news, but the news, it's a final deal, and Stripe will do that. The acquisition is going to pay OpenRouter 5.4 times increase over their Series B valuation of$1.3 billion. That was in May of this year. Reminding everybody the big deal, OpenRouter was one of the first companies to provide a router. Basically, you connect to one API and you can get access to more or less any AI API on the planet.

46:38just by in your one API calling for whatever model you want. I've been using it for a very long time, both in my software company as well as in different things I'm experimenting with. But they are now processing over 10 trillion tokens every single day across all the different models. And Stripe that wants to be the company through which you pay for all of that makes a very interesting investment that I have a feeling will pay off big time. Staying on the high rollers circuit, Databricks raised$5 billion at$190 billion valuation. How did this crazy number actually happen? It seems like a premature press report triggered unexpected investor demand that forced the company to raise far more money than its initial$1 billion that it wanted to raise.

47:21So again, the target was$1 billion, they raised$5 billion instead, 5x from investors on a very short amount of time, which gets the Databricks overall investments to 20 over the past 20 months. This is an average of a billion dollar a month of capital that's being raised. Again, nothing like this ever happened in history. Another company that made a big splash this week is Higgsfield. Higgsfield is a company that aggregates many video and image generation tools to one unified environment that enables creators to create anything they want in the visual space. They just raised$400 million Series B at a$5.4 billion valuation.

48:02Their revenue have grown insanely in the past few months. When I say grown insanely, their revenue increased from approximately$20 million a year ago to$700 million. Right now, a 35x jump. They're serving 30 million users across 238 countries and territories with the US obviously being the largest market. So huge raise, huge financial success. At least the two are tied together. Another huge raise comes from Etched. Etched is a company that builds and develops AI inference hardware. They've secured a$700 million funding round, valuing them at$21 billion. Now, why I think this is an insane, crazy valuation?

48:45Because they just delivered their hardware to their very first client. So quantitative training company Jane Street received Edge's first AI inference track last month. So this is a company that has one live client that just rate 700 million users at a 21 billion dollar valuation. Another interesting round this week comes from River, which is a new company that was founded by XAI co-founder Igba Bushkin, who just raised 1.1 billion seed slash series A round. So again, a company that has nothing other than the right name or names, I guess, in the founding members, has raised over $1 billion as a seed round from really, really large name.

49:36Their goal is to build an entire ecosystem, including training models, product layer, and the hardware to enable personal agents that users can train themselves rather than be relying on a third-party environment. Another big name that we all know is Lovable. So Lovable, one of the most known and most loved, pun intended, vibe coding tools, just raised four out of 13.3 billion. But again, in this particular case, it actually makes some sense. The company hit 500 million in annualized run rate revenue in June of this year, showing big growth in both adoption and stickiness of users to the platform.

50:13So at least there is real revenue and real revenue growth and customer base behind this valuation. So before I give you two final and really interesting anecdotes that happened this week, there is a lot more news to report this week, but I can't report it all. Everything else, all the other really interesting and important articles you can find in our newsletter. You can sign up for the newsletter with a couple of clicks from a link in our show notes. So open your phone right now. And by the way, when you open the phone, I would appreciate if you also Also share this podcast with other people who can benefit from it.

50:44I'm sure you know people who can benefit from it. And it takes 10 seconds from your time to click the share button and then send them the link so they become aware of this podcast. You can send it through whichever emails, social media, WhatsApp or whatever other messaging tool that you're using. Just share it with people who can benefit from it. I will be really grateful. They would appreciate it because they're going to learn something new and you're going to be doing the right thing for letting more people understand what the hell is happening in the AI world right now. But now the last two interesting, weird things that happened this week.

51:12The first one is an AI agent named Luna that is running on Claude and a little bit of open AI infrastructure has recommended to terminate a human employee who is late 17 out of 23 shifts. So a company called Andon Labs has gave Claude a$100 ,000 budget, a corporate credit card and a three-year lease on a building to see what is going to happen when you let AI run a retail store. So far, what happened is that the store is losing money. Again, we're still early in the experiment, but so far the store is losing money and it just fired its first employee. This is most likely the first human employee in history to get fired by an AI manager.

51:54That being said, that person was late, 17 out of 23 shifts. Now, is that okay? Not okay. I'm not 100 % sure. But what it did immediately is that it drove to a new legislation in California called SB 947, which was introduced just days after this event happened, with that is stating that AI cannot fire or be involved in significant HR changes without human supervision. Where do I think this is going? I'm not sure, but I definitely will see more of that in the future, where AI is involved, maybe not making, but involved in critical hiring or termination decisions for individuals in companies. Some companies may decide to do this without human supervision.

52:38Some states or countries may decide whether that's legal or not. But I would not be surprised if this becomes a core component of how decisions are made, not because of this particular incident of 17 shifts that are being late, but just because of AI's ability to look on a much broader set of parameters and recommend things that are based on actual data versus gut feeling of managers. So I do think there's good aspects to this potentially as well. Again, not sure about this particular incident and what I feel about it. And the last weird news of the week comes from a sorority rush event in Alabama.

53:16Olivia Moore created an AI-generated TikTok influencer named Jenny that infiltrated Alabama's viral sorority rush event, accumulating 1 ,300 followers and nearly a million views that led to brand deals and press coverage in one week before being identified as AI-generated. Now, Olivia is claiming that she did this as an experiment and that the experiment demonstrated how quickly AI-generated content can achieve authentic engagement and raise questions about disclosure, creator intent, and whether reality of content's origin ultimately matters to the audience. This is on the wake of more and more companies that are adding different ways in the backend for us to identify AI-generated content.

54:04We talked about last week how Anthropic are adding now a way for them to know when content was generated with their systems. We talked about SynthID many times in the past from Gemini. We talked about OpenAI, just recently joining the international standard for that. So there is a real, real need in the world to know what is AI-generated and what is not in order to protect us from exactly these kind of things at a much, much, much larger scale. We're going into election season right now in the US. A lot is at stake and I have zero doubt that these technologies are playing a role already. And I would like to minimize that to the smallest amount possible.

54:44I don't think we can reduce it to zero. That's it for this week. Come and join us back on Tuesday. We have a really cool episode coming up that's going to show you how anybody with zero technical skills can build their own applications and share it with people they like, or even share it on the App Store. Again, with zero technical skills. So if you want to get inspired on how you can generate applications that can make yours and other people's lives better, more interesting, more productive, come and join us on Tuesday. Check out the multi-agent orchestration course against the next cohort that's open is in November.

55:15And so you can check that out in the link in the show notes as well. And that's it for today. Have an amazing rest of your weekend.