Maybe AI Will Cure Cancer After All

0:00This podcast is supported by Google. Hey folks, Stephen Johnson here, co-founder of Notebook LM. As an author, I've always been obsessed with how software could help organize ideas and make connections. So we built Notebook LM as an AI-first tool for anyone trying to make sense of complex information. Upload your documents and Notebook LM instantly becomes your personal expert, uncovering insights and helping you brainstorm. Try it at notebooklm.google.com. Today on the AI Daily Brief, maybe we're going to get that AI cancer cure after all. And before that, in the headlines, Google announces VO 3.1, but how does it hang compared to Sora 2?

0:40The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:51All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, Gemini, Notion, Blitzy, Superintelligent, and robots and pencils. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. If you're interested in sponsoring the show, you can email us at sponsors at aidailybrief.ai. Two last quick announcements before we get into the episode. First, as I've mentioned over the last couple of days, I have a podcast growth roll-up. The role is simply to grow this show as big as humanly possible, and the way that you apply is doing something interesting to grab my attention.

1:24All of that can be found at aidailybrief.ai slash jobs. Lastly, I've had a super positive response to that episode from last Sunday, What 1000 Executives Say About AI Agents, basically the one where I got deep on what we've learned at Superintelligent. And so as a little bonus, I've decided to pull in Superintelligent's head of research, Nufar, who has been a guest on the show before, to do a three-part mini-series on trying to turn what we've learned into more practical, actionable lessons. So we've got an episode on culture, an episode on tech and data readiness and an episode on use cases. And I'm going to use the next three Saturday slots, which as you guys know, is historically the one off day for the week to have those as bonus episodes.

2:04So look for the first one of those on Saturday. With that though, let's get into today's episode. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. We kick off today with a model update that many people have anticipated that has been well hinted over the last few days. Google has released an iterative update of their video model, VO 3.1. The new version improves the realism of outputs, boosts prompt adherence, and improves audio quality. Still, the big change is new edit features. Users can now include reference images for objects and characters, as well as prompting the model to remove a particular object from a previously generated clip.

2:41They can also provide a first and last frame for a video and prompt the model to fill in the rest. Additionally, there's a feature to extend clips based on the last few frames, allowing creators to easily string together clips into minute-long shorts. The update comes five months after the May release of VO3, which was the single big game changer for AI Video, introducing synced audio for the first time as well as delivering state-of-the-art realism. That said, a lot has changed since VO3 was released, and while it sparked wonder and creativity, the general sentiment around 3.1 has been far less enthusiastic.

3:13AI developer Matt Schumer wrote, My initial VO3.1 impression? Disappointment. Unfortunately, it's not just noticeably worse than Sora 2, it's also quite a bit more expensive. One bright spot is the tooling they've added. It seems to me a little bit like the normal pattern, where the disappointment stems from the fact that this is just an iterative update, not some big state-of-the-art advance. VC Justine Moore noted that we've passed the threshold where video models are good enough, so we shouldn't expect anything new to be all that mind-blowing. She commented, We have entered the product era for video models.

3:44The recent releases, via 3.1, Sora 2, runway apps aren't a huge leap forward in terms of underlying model capabilities, but they introduce critical features like extending video, character consistency, and editing. I think that is a perfect summary of where things are. Expect that a lot of the updates in the immediate term when it comes to AI-generated video are going to be around how usable it is in production environments. Speaking of new models, Anthropic has released Claude Haiku 4.5, the latest version of their small model. The model is intended to be fast and cheap, with the claim being twice the speed of Sonnet 4 at a third of the cost.

4:18Anthropic also claims that the new version of Haiku outperforms the previous generation, Sonnet 4, in software engineering in the Sweebench verified test. They're also seeing outperformance against Sonnet 4 on computer use tasks, which could make the new Haiku a very capable agentic model. Anthropic Chief Product Officer Mike Krieger said, It's opening up entirely new categories of what's possible with AI in production environments, with Sonnet handling complex planning, while Haiku-powered sub-agents execute at speed. We're giving people a complete agent toolbox where each model has the right combination of intelligence, speed, and cost for different parts of the job.

4:52That's exactly what Kat Wu from Anthropic said. Haiku 4.5 is a workhorse that makes the coding experience in Cloud Code feel really fast. While Sonnet 4.5 remains the default, Haiku 4.5 now powers the Explore subagent, which can rapidly gather context on your codebase to build apps even faster. Haiku 4.5 will be available to free users and can be used to squeeze more capacity out of the free service compared to Sonnet 4.5. Krieger again commented, Even for my own use, even though it's not as smart as Sonnet, I've started defaulting to it on Claude, especially in the mobile app, because it's just much faster getting an answer.

5:26Putting the model through its paces, Swix was impressed, posting, More than twice the speed is underselling Haiku, to be honest. I built a way to directly compare Sonnet versus Haiku 4.5, and it's roughly 3.5 times faster, but the UX feels so much better because Haiku stays inside the flow window. Obviously, end-to-end latency varies a lot, so Anthropic can't report a real number without production usage, but you should try heads-up comparisons. One other quick note about Anthropic, we got some absolutely monster revenue numbers reported by Reuters from that company. Their sources suggest that Anthropic is currently running at a$7 billion run rate and is on track to hit$9 billion by the end of the year, and then get to somewhere between$20 to$26 billion next year.

6:06I'm not going to go too deep into that today because I'm planning to do a deep dive analysis on all of the implications of that and reported OpenAI numbers, probably for tomorrow's episode. But suffice it to say that Anthropics coding in an enterprise business is going very, very well right now. One company whose AI business is not going so well, to the extent that it can even be said to exist, is of course Apple. Another high-profile Apple AI researcher has left Apple to sign on with Meta's superintelligence team. Bloomberg's Mark Gurman reports that Qi Yang has left Apple just weeks after being promoted to lead the Answers Knowledge and Information team.

6:40That team was formed recently to develop a perplexity-style AI search product, which was viewed as a central pillar of a major Siri revamp planned for release in March. Yang was one of the most senior executives among Apple's broader AI and machine learning group. Gurman wrote that this is one of the most high-profile exits from Apple's AI organization, which has seen about a dozen departures this year. What's more, his sources said that even more departures are expected over the coming months. Gurman concluded, The continued departures underscore the instability within Apple's AI ranks at a time when it's racing to catch up with OpenAI and Google, both of which are advancing quickly in generative AI and search.

7:14He noted that Apple had also been interviewing outside replacements for John Gianandrea, Apple's senior VP of AI and machine learning, who leads the entire AI organization. Lastly today, a study which I found to be just a real bummer as an American citizen, Pew Research has published the results of a new survey showing that global public sentiment is souring on AI. They interviewed people across 25 countries and found that in general they are far more concerned than excited about the increased use of artificial intelligence. Overall, 34 % of respondents said that they were more concerned than excited, while only 16 % said that they were more excited than concerned.

7:4842 % said that they were equal parts excited and concerned. Across all 25 countries, there was not a single one where excitement was the majority feeling about accelerated AI adoption. I should note here that the big notable exception to countries that were included, there is no China here. And I would be very interested to see what those numbers look like. Still, of the countries that were surveyed, only three had more than 20 % of their people saying that they were more excited than concerned. Nigeria was at 20%, Sweden and Korea had 22 % each, and Israel was at 29%. Israel and South Korea were in fact the only nations where the people who said they were mostly excited about AI outweighed the people who said they were mostly concerned.

8:27The part that I said was disappointing to me was that right at the very top of the list for populations most concerned about AI was the US at 50%. While technically we tied with Italy on that number, Italy had more people who were more concerned than excited than us, 12 % compared to just the one-tenth of Americans who are more excited than concerned. Now, it should be noted that this data is a little old at this point. Two US surveys were conducted in March and June, while the international survey was conducted between January and April. But given that we've seen the rise of protests around data center construction and a lot of negative news reporting on AI, I would be surprised if we had seen a major improvement in sentiment, and I wouldn't be surprised if we'd seen it actually get worse.

9:07Now what's crazy about this is that you have significant portions of these populations actually using these tools, and yet they still have this anxiety. There is a lot more work to be done for those of us inside the AI industry, not only from a PR type of perspective, but in ensuring that people's concerns about their futures are actually addressed. My general thesis is that AI is the recipient of more generalized anxiety and that everything is downstream from economic insecurity. So who knows, one day maybe I'll have to spit out the politics podcast. For now, I'll just leave it at more work to be done, and that's going to do it for the headlines.

9:39Next up, the main episode.

9:44Chatbots are great, but they can only take you so far. I've recently been testing Notion's new AI agents, and they are a very different type of experience. These are agents that actually complete entire workflows for you in your style. And best of all, they work in a channel that you already know and love because they are purpose-built Notion super users. Notion's new AI agents completely expands the range of what Notion can do. It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days, and get them complete in minutes.

10:16These agents don't just help with work, they finish it. Getting started with building on Notion is easier than ever. Notion agents are now your very own super user to help you onboard in minutes. Your AI teammates are ready to work. Try Notion AI for free at the link in our show notes. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Enterprise engineering leaders start every development sprint with the Blitzy platform, bringing in their development requirements.

10:50The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Blitzy is providing a limited-time, 30-day free proof of concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org.

11:23Visit blitzy.com and press book demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native. That's blitzy.com. Today's episode is brought to you by my company, Superintelligent. You've got 100 what-if ideas, but which one becomes an agent? Superintelligent maps every AI use case across your company and helps you create an agent plan that you can actually execute. We match opportunities to your tech stack, your data profile, and your team. No more guesswork, just a clear path from pilot to production. If you want agents that deliver business outcomes, start with planning. Go to bsuper.ai and sign up for a demo.

12:03AI isn't a one-off project. It's a partnership that has to evolve as the technology does. Robots and pencils work side-by-side with clients to bring practical AI into every phase. Automation, personalization, decision support, and optimization. They prove what works through applied experimentation and build systems that amplify human potential. Welcome back to the AI Daily Brief. The joke over the last couple of weeks, first as OpenAI launched Sora, a short-form video app, and then later as it announced that it would be opening up to adult uses of its platform, goes along the lines of, we were promised great big cures for diseases and novel discoveries, and instead, we got a new TikTok and a new place for porn.

12:47Major politicians even weighed in on this line of memeing, with Florida Governor Ron DeSantis saying so much for curing cancer and beating China? And yet, even as that discourse was happening, we got this announcement from Google's Sundar Pichai. He writes, An exciting milestone for AI and science. Our C2S-scale 27B foundation model, built with Yale and based on Gemma, generated a novel hypothesis about cancer cellular behavior, which scientists experimentally validated in living cells. With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer.

13:24So today we're going to talk about first this particular discovery, and then more broadly how quietly, behind all the hype and noise, there have been some really interesting advancements which suggest that the whole idea that AI can't make or contribute to novel discoveries in science may be one now for the junk heap of history. So back to this discovery from Google. In their announcement post, Google wrote, This announcement marks a milestone for AI in science. C2S scale generated a novel hypothesis about cancer cellular behavior, and we have since confirmed its prediction with experimental validation in living cells.

13:58The implications, they say, are new pathways for developing therapies to fight cancer. Now, they explain that one of the biggest challenges in cancer therapy is that many tumors are quote-unquote cold. In other words, invisible to the body's immune system. A major strategy in cancer treatment, then, is triggering tumorous cells to make them turn hot, i.e. to display immune-triggering signals, in a process that's called antigen presentation. With this as background, researchers gave C2S scale a single task, to find a drug that functions as a conditional amplifier. In other words, to boost the immune signal only in specific circumstances.

14:34Previous iterations of similar models were not capable of achieving this task, but C2S scale succeeded. The task effectively required a sort of conditional biological reasoning. They designed what they called a dual-context virtual screen, where they, one, provided the model with real-world patient samples with intact tumor-immune interactions and low-level interferon signaling, and then secondly provided the model with isolated cell line data with no immune context. Google then simulated the effects of over 4 ,000 drugs and asked the model to predict which would boost immune signals if only certain conditions were met.

15:08Now, this highlights one of the areas where we're seeing AI-enhanced science really flourish. AI models generally excel in situations where a large volume of experimentation is required. In other words, a big part of the value is about speeding through simulated experiments and crunching large data sets that would take human researchers and traditional computing methods much, much longer to sift through. After simulating those 4 ,000 drugs, the experiment found a set of drug candidates. Out of the drug candidates that the model highlighted, only 10 to 30 % were already known in prior literature.

15:40The others had no prior link to the screen. Interestingly, the model made a core prediction on how the family of drugs would function, which it used to base its result. They wrote, What made this prediction so exciting was that it was a novel idea. The model was generating a new testable hypothesis and not just repeating known facts. Researchers then tested the hypothesis on actual cells and observed the predicted effect. The model seems to have correctly identified a new way of turning tumorous cells hot under the desired conditions. Google concluded, While this is an early first step, it provides a powerful experimentally validated lead for developing new combination therapies, which use multiple drugs in concert to achieve a more robust effect.

16:19This result also provided a blueprint for a new kind of biological discovery. It demonstrates that by following the scaling laws and building larger models like C2S scale 27b, we can create predictive models of cellular behavior that are powerful enough to run high-throughput virtual screens, discover context-conditioned biology, and generate biologically grounded hypotheses. One of the big implications here is that these larger science-specific models seem to actually have emergent capabilities in scientific reasoning, not just language-based reasoning. To the extent that this is a bitter lesson outcome, i.e.

16:51just the byproduct of a better, bigger, more dedicated model, that actually makes it more likely that this is a big unlock for future research rather than a one-off discovery. Basically, the implications of there being a general scaling law for scientific reasoning models is quite large. The reactions from many were excited. We got, of course, the jokes. Packing McCormick wrote, everyone else, behold, an AI you can beat off to. Google DeepMind, protein folding, weather prediction, new materials, and now an AI that can make its own cancer discoveries. There was, however, some skepticism. Lenny Eusebi writes, A bit of a stretch to frame this is if they asked a chatbot to solve cancer and it spat out a novel idea.

17:27This is much more like they trained a narrow predictive model for a very specific task, and then it was able to do that task well enough to filter out a new candidate drug. Some version of this take is basically presented in every thread. The point, though, with this discovery is that the model demonstrated the ability to take a set of known facts about the science and synthesize them into a novel hypothesis that proved to be correct using reasoning. If you go look at these threads where inevitably this critique comes out, there are scientists who follow up, pointing out that there's really no such thing as scientific discovery created from whole cloth.

17:57Everything is built on the synthesis of existing ideas. Rob S. follows Lenny's post with, yes, that's how science is done. VC Hemant Mahoptra writes, I've always believed new knowledge can be, one, built on existing knowledge but connecting the dots in unique ways, two, creating pure de novo knowledge through hypothesis experiments, etc. that might go against current thinking. LLMs are likely graded one, and that's where perhaps a vast majority of the net new knowledge lies. Even if LLMs as they stand today never get to number two, their impact on research will be tremendous. Now, what makes this story notable to me, even outside just the profound implications of AI actually being able to help us cure cancer, is that it is not an isolated story.

18:37For those who have been paying attention closely, and of course that's hard considering the absolute barrage of new models and crazy bubble talk and all those things going on, there have been a lot of these really subtle indicators that some big barrier has been surpassed. OpenAI's Kevin Wheel, who used to be their chief product officer but is now their VP of science, about a week ago tweeted, GPT-5 crossed a major threshold. Over the last two months, we've heard repeated examples of scientists successfully directing GPT-5 to do novel research in math, physics, biology, computer science, and more.

19:07Now he clarified, I'm not claiming GPT-5 is ready to prove the Riemann hypothesis. It's more at the lemma stage today when guided by an expert, it can do bounded chunks of novel science. Things that would maybe have taken a professor on her postdoc a few days or a week to work through. But this is the beginning of accelerating science, because if each path takes a week, you can only explore so many of them. If it takes 20 minutes with ChatGPT Pro and you can run them in parallel, suddenly you can explore far more. And remember, the model you're using today is the worst it'll ever be for the rest of your life.

19:36The idea that ChatGPT could do novel science sounded crazy a year ago, but here we are. And by the way, this is not just Kevin speaking. Professor Ethan Malik wrote, I'm hearing similar things in economics and the social sciences. Not autonomous work, but expert-directed AI is absolutely helping academics do novel research in significant ways. One example that got a lot of conversation came from back in August. Sebastian Bubeck, a researcher at OpenAI, posted an academic mathematics problem to GPT-5, and it appeared to come up with a novel result. The problem was an extension of existing work, which Bubeck explained as, in smooth convex optimization, under what conditions, on the step size eta and gradient descent, will the curve traced by the function value of the iterates be convex?

20:17It's totally fine if that's gibberish to you, it is absolutely gibberish to me. The fact that you need to understand is that the original paper on Arvix found a general result if the eta is larger than 1.75 divided by L, where L is the smoothness of the curve. The paper also provided the result below 1 divided by L, so there was a remaining gap between 1 and 1.75. GPT-5 Pro appeared to produce a general result for 1.5 divided by L, reducing the lower bound of the solution. Bubeck commented that this was, quote, definitely a novel contribution that'd be worthy of a nice Arvix note. However, he continued, the only reason I won't post this as an Arvix note is that the humans actually beat GPT-5 to the punch.

20:54Namely, the Arvix paper has a V2 with an additional author, and they closed the gap completely, showing that 1.75 L is the tight bound. Still, he pointed out that GPT-5's proof was completely novel, commenting, The fact that it proves 1.5 divided by L and not the 1.75 divided by L proof also shows that it didn't just search for the V2. Also, GPT-5's proof is very different from the V2 proof. It's more of an evolution of the V1 proof. Shortly after Bubeck published his results, others at OpenAI chimed in that this wasn't the only Alval academic work that GPT-5 was capable of. Chief Research Officer Mark Chen posted, GPT-5 Pro is starting to develop new mathematics.

21:29I'm hearing similar stories in other scientific domains like physics too. Now what's interesting about these math results is that as much as we are talking about AI's ability to generate new knowledge by synthesizing old knowledge as a pathway for medical and scientific discovery, this math result seems to be an emergent capability of reasoning models. In coming up with the proof, GPT-5 Pro thought for 17 minutes and then presented work that wasn't previously published. Then again, maybe we shouldn't be so surprised given recent performance. Both OpenAI and DeepMind entered LLMs in the International Math Olympiad this summer and were capable of gold medal performances.

22:02The notable thing is that these kinds of theoretical math problems have basically zero calculation. They're about manipulation of logic to come up with a mathematical proof. It's basically an entirely different category of scientific work. Former quant investor Jeffrey Emanuel highlighted another interesting novel math paper that required a lot of manual labor to come up with the result. In a long thread, he suggested that this could be an example of a hidden discovery. A novel result that was already feasible based on current knowledge, but required too much work for a human to reasonably obtain as an individual or an academic team.