When Will AI Make Scientific Discoveries?

0:00Today on the AI Daily Brief, when will AI start making novel scientific discoveries? Before that, in the headlines, the US government says actually DeepSeek stinks. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:21All right, quick announcements before we dive in. First of all, thank you to today's sponsors, Notion, Blitzy, InsightWise, and Pencils & Robots. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can now subscribe to the audio-only version ad-free in Apple Podcasts. So if you just listen and you get it through Apple Podcasts, you can now subscribe directly there. I'm working on having Spotify's ad-free subscription coming as well. So soon you will have multiple choices, Patreon, Apple, and Spotify. But with that, let's dive in. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes.

0:55We kick off today with a new report from the National Institute of Standards and Technology that finds DeepSeq lacking in both performance and security. Announcing the report, Commerce Secretary Howard Lutnick wrote, Today, thanks to President Trump's AI action plan, the Commerce Department and NIST's Center for AI Standards and Innovation have released a groundbreaking evaluation of American versus adversary AI. Result? American AI models dominate. Our systems outperform DeepSeq across nearly every benchmark. The report is clear. DeepSeq lags far behind, especially in cyber and software engineering.

1:30These weaknesses aren't just technical, they demonstrate why relying on foreign AI is dangerous and short-sighted. Allowing our adversaries to control AI poses serious risks to our security. By setting the standards, driving innovation, and keeping America secure, the Department of Commerce is helping ensure continued U.S. leadership in AI. Now, in addition to lack of performance, NIST found that DeepSeek was more expensive than comparable U.S. models. They claimed that one of the U.S. models used as a reference cost 35 % less on average to complete the 13 performance benchmarks that were tested.

1:59NIST also found that DeepSeq's models were 12 times more likely than U.S. frontier models to allow malicious instructions designed to derail them from the desired task. NIST wrote, hijacked agents sent phishing emails, downloaded and ran malware, and exfiltrated user login credentials all in a simulated environment. DeepSeq's models were also far more susceptible to jailbreaking, responding to 94 % of malicious requests after using a common jailbreaking technique compared to just 8 % for U.S. reference models. NISC found that DeepSeq's models echoed four times as many inaccurate and misleading CCP narratives as US models.

2:31And finally, and perhaps most concerningly, they found adoption of Chinese models has greatly increased since DeepSeq R1 was released. They anchored this claim on the statistics that downloads of DeepSeq models are up 1000 % since January of this year. Now, some people were a little skeptical of this, but mostly it was just a Rorschach test for wherever your politics happened to already be. How much it changes any sort of AI policy, I'm a little bit more skeptical of. Next up, a little in the AI hardware realm, Apple has scrapped plans to iterate on the Vision Pro and will shift focus to developing AI smart glasses.

3:01It seems that Meta's Ray-Bans were just too compelling, forcing Apple to scuttle their existing plans. Bloomberg's Apple specialist Mark Gurman reports that Apple had plans to develop a cheaper, lightweight version of the Vision Pro, which was on track for release in 2027. Sources said that last week, however, Apple announced internally that staff would be pulled off that project to accelerate work on smart glasses instead. Plans are reportedly to work on two different versions of the glasses product. A cheaper version dubbed the N50 will compete with the original Meta Ray-Bans, and a higher spec version will include a display to go up against the newly released Meta Ray-Ban display.

3:33The N50 is expected to be ready for unveiling next year ahead of a 2027 release, while the display version isn't expected until 2028. The design for Apple's glasses seems to stick close so far to Meta's product line. Apple plans to use voice controls and integrated AI as the core interface. The glasses will also feature speakers for music and cameras for media recording. And Apple is also reportedly exploring integrating health tracking capabilities for the device as well. The reporting also seems to suggest that Apple has come to consider it a mistake to have tried to sell a$3 ,500 consumer device that's not comfortable enough to use for long periods of time.

4:05Right, Skerman, Apple executives have acknowledged the product shortcomings in private, viewing it as an over-engineered piece of technology. The game's not over, but it certainly puts more evidence of the column that smart glasses, for the moment, are the very tippy-top of the format war for AI devices. Speaking of devices, Amazon has released a new line of smart devices built specifically for their AI assistant Alexa Plus. The entire range of Echo smart speakers have been upgraded with new custom silicone featuring an AI accelerator. This will allow the new devices to provide local inference for Amazon's AI models, and the devices will also include a new custom sensor platform designed to make ambient AI feel more natural.

4:42The sensors include cameras, audio, ultrasound, Wi-Fi radar, and an accelerometer. Basically everything imaginable to make the AI model aware of its surroundings. The goal is to make interactions with the ambient AI feel more natural and responsive. Some of the devices enable new AI features across Amazon's smart home systems. The latest Ring cameras now have facial recognition technology that can keep track of friends and family and distinguish them from lurking strangers. Another feature called Search Party allows users to send out an alert across networked cameras in a neighborhood to search for a lost pet.

5:10This is the first big product refresh since Amazon hired product chief Panos Panay away from Microsoft back in 2023. Panay told Bloomberg, my belief is that our job is to make devices the next big business at Amazon. And AI is very clearly right at the core of the strategy, which Panay articulated as, great products made even better through ambient AI. Lastly today, a relevant one, given all the dust up around Sora 2 and the presumption that all of this is leading to ads in ChatGPT, Meta has crossed the Rubicon and will start to target ads based on users' AI chats. Meta announced a change to their recommendation system on Wednesday that will see AI interactions used to personalize content and advertising delivery across their apps.

5:50Meta gave the example of a user asking their chatbot for nearby hiking recommendations. The user might then be served hiking-related content and ads for hiking boots or other gear. Users won't be able to opt out, but sensitive topics will be automatically excluded. These include politics, religion, sexual orientation, health, and racial origin. The change will go live in December, and the policy won't apply in the UK, Europe, and South Korea due to their stricter tech privacy rules, although Meta plans a compliant rollout at a later date. Christy Harris, the privacy policy manager at Meta, said, People's interactions simply are going to be another piece of the input that will inform the personalization of feeds and ads.

6:23We're still in the process of building the first offerings that will make use of this data. Look, it feels pretty inevitable that ads are going to be a part of the AI landscape. I think to some extent, the question will be for people who are already paying for subscriptions, are they going to have to deal with ads as well? What sort of privacy controls will there be? All of these questions remain to be seen, but it is not surprising that we are getting to this point. Frankly, it's just surprising that it's taken this long. For now, that's going to do it for today's headlines. Next up, the main episode.

6:50Chatbots 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.

7:22These 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 superuser 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.

7:56The 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.

8:28Visit blitzy.com and press book demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native. That's blitzy.com. As a consultant, responding to proposals can often feel like playing tennis against a wall. You're serving against yourself trying to guess what the client really wants. That all changes with the InsightWise proposals platform. Now you've got an AI coach that thinks just like your client. It returns to the brief time and time again, identifying opportunities, showcasing your track record, and making recommendations to improve your pitch. Suddenly you're on center court, but this time you've got a secret weapon.

9:04InsightWise does away with all the time-consuming manual work, so you can focus on winning more business more often. Generate reports, pull insights from your own data, build competitive advantage, and go to sleep before 2am. When it comes to proposals, you only get one shot. With InsightWise, make yours an ace. AI 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.

9:41As an AWS-certified partner with Global Delivery Centers, Robots and Pencils combines reach with high-touch service. Where others hand off, they stay engaged. Because partnership isn't a project plan. It's a commitment. As AI advances, so will their solutions. That's long-term value. Progress starts with the right partner. Start with robots and pencils at robotsandpencils.com slash AI Daily Brief. Welcome back to the AI Daily Brief. Today we are discussing AI and scientific discovery. And of course, at least part of the context has to be the launch of Sora 2. We talked extensively about this in yesterday's episode, but if you had to take just one meme that best summed up the enfranchised critique, let's say, of OpenAI's announcement, not just of the Sora 2 model, but of the Sora app to go with it, it was really a critique of what OpenAI is choosing to spend its time on.

10:31One example of this was Rudder Tushar who said, Sam Altman two weeks ago, we need$7 trillion and 10 gigawatts to cure cancer. Sam Altman today, we are launching AI slot videos marketed as personalized ads. Now, there was a lot of that going around, enough that it clearly got under Sam Altman's skin. In fact, he responded to that one saying, I get the vibe here, but we do mostly need the capital for building AI that can do science. And for sure, we are focused on AGI with almost all of our research effort. It is also nice to show people new cool tech products along the way, make them smile and hopefully make some money given all that compute need.

11:05When we launched ChatGPT, there was a lot of who needs this and where is AGI? Reality is nuanced when it comes to optimal trajectories for a company. So let's hold aside Sam's response and whether you think it's legitimate or not. The point is that a lot of people are saying, we were promised AGI and scientific discovery and we got another social media app. It harkens back to the famous Founders Fund manifesto, where they wrote, we wanted flying cars, instead we got 140 characters. This sense that there is some deeper value to big, massive machines and inventions and discoveries, as opposed to just social media attention sucks.

11:38This question has actually been lurking around the open AI space for a lot of 2025. One of the big stories of this year, of course, was Mark Zuckerberg parading around the valley, trying to poach engineers to stock his super intelligence lab. And while a lot of that effort was very successful and a lot of incredibly talented people came over to Meta, there were some who turned down even extremely generous compensation packages. And a lot of the scuttlebutt and buzz around AI circles was that those folks just weren't willing to risk that ultimately what they were going to have to spend all of their brainpower and all of Meta's compute on was improving ad click-through rates.

12:11Now, one of the things that was interesting about this conversation yesterday, as it happened alongside Sora, was that we didn't just have to speculate that some number of researchers and top minds weren't going to be interested in that sort of work. We actually had an example of researchers who had left OpenAI and Google and Meta to build something that had a much heavier scientific discovery type of focus. The company was called Periodic Labs. Its goal is to use and build AI that can actually accelerate discovery in fields beyond computer science, such as physics and chemistry. And its announcement was very explicitly positioned as being about AI researchers getting sick of working on consumer AI and moving on to a higher purpose.

12:54Indeed, the New York Times article about Periodic Labs starts with one of these stories. They write, This summer, Mark Zuckerberg invited Rashab Agrawal to join the company's new AI lab, offering him millions of dollars in stock and salary. With the new lab, Zuckerberg said, he wanted to build superintelligence, a technology that could eclipse the power of the human brain. Though no one knew how to create superintelligence, he urged Dr. Agarwal to make a leap of faith. In a world that is changing fast, Zuckerberg told him, the biggest risk you can take is not taking any risk. But although Dr. Agarwal was already a meta-employee, he turned down the offer to join another company.

13:26That company is, of course, Periodic Labs. In discussing Periodic's goals, founder Liam Fita said, the main objective of AI is not to automate white-collar work. The main objective is to accelerate science. Now, we talk in this show all the time about the difference between efficiency AI and opportunity AI. And this is, of course, in the context of enterprises deploying AI at work. The idea of efficiency AI is thinking about AI simply as a way to do what is currently done, but faster, cheaper, or maybe better, but still doing the same thing. There's nothing wrong with efficiency AI. People should leverage those gains.

14:00They're going to become table stakes. But the real opportunity and what will differentiate companies I have always said and believed is those who think about it as a new opportunity technology, a technology, in other words, that opens up things that weren't possible before. The founders of Periodic Labs are taking a similar assessment and have built their company around the premise of answering how to actually make that real and figure out a gap in the market that opens up that possibility. So what does Periodic Labs do? Simply put, their goal is to accelerate science. In their announcement post, they wrote, Our goal is to create an AI scientist.

14:32Science works by conjecturing how the world might be, running experiments, and learning from the results. Intelligence is necessary but not sufficient. New knowledge is created when ideas are found to be consistent with reality. And so at Periodic, we are building AI scientists and the autonomous laboratories for them to operate. And really at core of what they think is missing is that while yes, current models have read everything that's available, ultimately to make new discoveries you need practical application and experimentation. As they put it, as any scientist knows, though rereading a textbook may give new insights, they eventually need to try their idea to see if it holds.

15:06So basically, what they want to do is connect the dots between human researchers, AI agent experiment designers, and autonomous and robotic labs where those experiments can be conducted. The way the neuron framed it was this. Human evaluers initiate, AI agent designs experiments, the robotic lab executes them, and nature itself provides a reward signal, did the experiment work, and from there data improves the models. Basically taking the scientific method and AI-ifying it. The company is starting with physical sciences, because they say physics is a verifiable environment. They note that AI has progressed fastest in domains with data and verifiable results.

15:42And this is what they mean when they say that nature is the reinforcement learning environment. Now part of the strategy is to collaborate with industry right from the get-go. For example, they are already working with a semiconductor manufacturer on issues around heat dissipation on their chips. Now, as part of the coming out party this week, Periodic Labs announced that it had raised over$300 million in seed funding from a who's who of investors, including Andreessen Horowitz, Excel, NVIDIA, Jeff Bezos, and many, many more. One of their investors, Bain Capital Ventures, wrote the rare investment announcement post that uses historical analogy well.

16:14They begin with an exploration of the difference in science before the telescope and after the telescope. They write, The history of science is full of similar examples in which technological progress enables the invention of new scientific instruments, which in turn leads to new scientific discovery. Basically that there is a relationship between the discoveries of science and the technology of science, that they are mutually reinforcing in a positive feedback cycle. The point is this. Galileo, they write, had the newly invented telescope. We have newly developed AI systems. What can we see now that we couldn't before?

16:45To the extent there was concern around the slopification of AI with Sora and before it, the MetaVibes app, the periodic announcement saw almost the inverse excitement, with just an incredible amount of enthusiasm, not just from the AI industry, but from many different parts of academia, science, research, and beyond. Now, one of the other comparison points in that August article about turning down Zuckerberg's offers was all of the people who were leaving to work with former OpenAI CTO Mira Mirati at her new startup, Thinking Machines Labs. And while Thinking Machines isn't as aggressively self-styled as a place for physical scientific research as periodic labs, they are very self-consciously trying to take a different approach to spreading and expanding knowledge around AI as opposed to the other labs.

17:25And right as OpenAI was announcing Sora 2, we also got the first product released from them. The product is called Tinker, and it's an API for training and fine-tuning custom models. Essentially, it's an AI infrastructure as a service, and it's meant to reduce the barrier to entry for model training substantially. Thinking Machines Lab provides the GPU cluster and the software stack, leaving customers to focus on training data and model design. Maradi posted, Tinker brings frontier tools to researchers, offering clean abstractions for writing experiments and training pipelines while handling distributed training complexity.

17:55It enables novel research, custom models, and solid baselines. In comments to Wired, she added, We believe Tinker will help empower researchers and developers to experiment with models and will make frontier capabilities much more accessible to all people. We're making what is otherwise a frontier capability accessible to all, and that is completely game-changing. There are a ton of smart people out there, and we need as many smart people as possible to do frontier AI research. Wrote Thinking Machine scientist John Shulman, Tinker provides an abstraction layer that is the right one for post-training R &D.

18:24It's the infrastructure I've always wanted. So the goal here is very much a democratization of frontier AI research. They're looking to help speed up innovation by enabling researchers or startups to test their ideas in days instead of weeks or months. They're trying to level the playing field, making it possible for smaller labs, universities, or even individuals to do meaningful AI work without billion-dollar budgets. They're looking for paths to make AI models more useful while keeping costs down. And certainly while there are powerful benefits from a broad research and democratization standpoint, there is also a ton to be excited about here for enterprises.

18:56In the first wave right after ChatGPT, a lot of enterprises tried to train their own models, mostly to discover the bitter lesson that generalist models simply outperformed their limited amounts of data, even if it was data that was contextual to their firm. Bloomberg was one big example of this. However, there's still a lot of interest in custom-trained layers for models that take advantage of proprietary and non-public data, and Tinker opens up some great possibilities on that front. Basically, if it works as promised, they will be able to fine-tune models on their corpus of data without big infrastructure and an extensive AI engineering team.

19:28What's more, it opens up the possibility that individual teams within an enterprise could actually think about this type of custom training rather than just having to wait for what the central AI group does. There's no reason, for example, that a marketing analytics team, especially if they had the right support, couldn't actually go explore that kind of customization, again, without having to get in line and wait for whatever other types of big AI infrastructure projects are going on across the company. Now, how much Thinking Machines is going to care about that enterprise use case remains to be seen, but it's certainly valuable.

19:56The first reactions are extremely positive. Technium, the co-founder of distributed training collective NOS Research, wrote, I had the privilege of being part of the beta for Tinker. It's a really nice project. Simple APIs for training models can make it a lot easier to properly leverage resources to get results. UC Berkeley PhD student Tyler Griggs writes, very hackable and lifts a lot of the LLM training burden, a great fit for researchers who want to focus on algs and data, not infra. Former OpenAI co-founder and VibeCode terminology coiner Andre Karpathy writes, Tinker is cool. If you're a researcher or developer, Tinker dramatically simplifies LLM post-training.

20:28You retain 90 % of algorithmic creative control, while Tinker handles the hard parts that you usually want to touch much less often, meaning you can do these at well below 10 % of typical complexity involved. Compared to the more common and existing paradigm of upload your data will post-train your LLM, this is, in my opinion, a more clever place to slice up the complexity of post-training, both delegating the heavy lifting, but also keeping majority of the data and algorithmic creative control. GDP at Amazon actually drew the connection between thinking machines and periodic labs. They write, thinking machines vision, whole world as an AIRL powered lab, In that sense, it's similar to that of periodic labs, but possibly much more expansive.

21:03Periodic labs will be the equivalent of Bell Lab of today's world, and they will conduct reinforcement learning on the basis of feedback on real-world feedback, e.g. experiments in material science. We need that, and I am rooting for their success. Thinking Machines aims to treat the whole world as an AIRL-powered lab. Currently, AI training, and particularly RL, is out of reach of most engineers and organizations. It is considered achievable only for a handful of labs with big egos. Also, big labs cannot collect real-world feedback data beyond a point. The real world provides the best reinforcement learning rollout data.

21:33For example, the kind of feedback Cursor gets when users either accept or reject suggestions by Cursor tab model. Customers support conversations and action trajectories that leave users delighted or disappointed. Factory floor decision-making data with outcomes. With such diverse data not accessible to the big labs as it is not on the internet, the models can be made much more intelligent. There is actually an entire show that I have had half formed on the back burner about the whole world as a reinforcement learning lab and how much the labs are orienting towards that being the future of model development.