Weekly Digest 38

Topics we are tracking

Anthropic Operates a Wet Lab in the Bay Area

Source: Reuters — Anthropic quietly sets up biology lab as it ramps AI drug program

Demis Hassabis, Sam Altman and Dario Amodei have all openly spoken about a future where AI could dramatically accelerate scientific progress and help cure some of today's most devastating diseases. DeepMind's AlphaFold was an early indication of this potential, although its impact so far has been much clearer scientifically than commercially.

Anthropic has also invested heavily in life sciences. Earlier this year it acquired Coefficient Bio, an AI biotech startup, and this week launched its Life Sciences Verification Program, which gives vetted researchers access to its most advanced models with fewer restrictions on biology-related work.

Anthropic is now also operating a wet lab in the Bay Area where it can use its AI models to conduct physical biology experiments. The company has declined to give specific information about what the lab is working on, but said that the lab is not specifically for drug discovery. That distinction may also help reassure pharmaceutical companies that are increasingly becoming its customers. Just this week it announced a partnership with Novo Nordisk to use Claude for drug discovery and development.

Combining AI models with physical labs is showing some interesting early results. Earlier this month Periodic Labs, which operates high-throughput materials science labs in Menlo Park, introduced Periodic Neon, a model trained using data generated in its own labs. On an internal benchmark for analysing difficult X-ray diffraction measurements, Periodic says Neon outperformed GPT-6 Astra and Claude Fable 5.1 despite starting from an open-weight model.

Instead of training scientific models largely on existing internet and research data, companies can generate new data through physical experiments, train models on the results and then use those models to decide which experiments to run next. As the models improve, the labs generate better data, which can then be used to improve the models further.

Ultimately, we expect drug discovery to generate most of the commercial returns. It will be interesting to see how drugmakers like Novo Nordisk and Eli Lilly partner and compete with leading AI labs.

SK hynix Considers Bringing Memory Production to the US

Source: TrendForce

SK hynix is reportedly in talks with Intel to make memory chips in the US for the first time. It could lease part of Intel’s long-delayed Ohio factories or form a joint venture with Intel and large cloud companies seeking a reliable memory supply. Talks are still at an early stage, and the type of memory to be made there remains unclear.

SK hynix is already building a more than $4 billion HBM facility in Indiana, which broke ground last month and is expected to begin mass production in the second half of 2029. However, that facility is focused on advanced packaging. The DRAM wafers will still be manufactured in South Korea before being shipped to Indiana for stacking, packaging and testing. Manufacturing DRAM or HBM wafers in Ohio would therefore be a much bigger step towards building an end-to-end US memory supply chain.

The deal would provide Intel with another potential use for its Ohio fabs. Intel originally said investment in the site could eventually reach $100 billion and expected production to begin in 2025, but the first two fabs have since been delayed until the beginning of the next decade. Leasing part of the site to the world's largest HBM supplier could help Intel monetize some of that manufacturing footprint without having to fill the capacity entirely with its own products or foundry customers.

The relationship between Washington and Seoul has become increasingly strained in recent months. The US Commerce Department has repeatedly discussed higher tariffs on semiconductors, while Commerce Secretary Howard Lutnick has threatened chipmakers with tariffs if they do not move more production to the US. Seoul, meanwhile, considers advanced DRAM and HBM technology strategically important. Any transfer involving technologies classified as a "national core technology" would be subject to government review. HBM is one of South Korea's strongest bargaining chips in trade negotiations, so Seoul is likely to scrutinize any transfer closely. Semiconductor manufacturing also remains considerably more expensive in the US than in South Korea.

Personal AI Agents Start to Take Off

Source: TechCrunch

Meta launched its new personal AI agent Muse less than two weeks ago and early adoption has been strong. The app reached more than 730,000 US downloads during its first five days, slightly ahead of the Meta AI app over the same period, and briefly reached No. 2 in the US App Store.

As with xAI's Grok Bot, every Muse user gets a dedicated virtual computer with its own browser. The agent can connect to services like email, calendars and payments, remember information about the user and then take actions across different applications. Instead of asking an AI which hotel to book, the idea is that you tell Muse to organize the trip and it does the work itself.

Instinct, another startup focused on this sector, launched earlier this year and has already grown to more than 100,000 users. The company raised $250 million at a $2.5 billion valuation last month, bringing its total funding to $350 million. It is now reportedly discussing another $1 billion round at a valuation of around $10 billion. Instinct takes an even simpler approach: there is effectively no new interface. Users text or call their agent like they would another person, while the agent can access email, messages, screen activity, audio and location to complete tasks in the background.

Despite the different interfaces, all three products are moving towards a similar architecture: a persistent agent with its own computer, long-term context and access to the applications people already use. Claude Code and Work, as well as Codex, have so far been the primary interfaces for agentic AI for coders and knowledge workers. These new persistent "assistants" extend a similar experience to a much broader audience and could be the first real interaction with agentic AI for most consumers.

When OpenAI hired OpenClaw founder Peter Steinberger, he said he would work on bringing agents to everyone. Instead, both OpenAI and Anthropic have so far approached agentic AI primarily through coding and knowledge work, where agents operate in more structured environments, are easier to test and supervise, and have clearer economics. There may also be a more strategic reason: coding agents are increasingly being used to accelerate AI research itself, creating a feedback loop where better agents help build better models. If OpenAI and Anthropic believe this leads more quickly to much more capable systems, then building the perfect personal assistant today may simply be less important.

Meta and Mark Zuckerberg have been advocating for "Personal Superintelligence" for some time, and Muse is perhaps the first real iteration of that vision. Meta already has billions of users and direct access to their messaging and social applications, as well as one of the largest advertising ecosystems in the world, which could eventually provide a natural path into commerce and transactions through these agents. Grok Bot similarly has an existing distribution channel through Grok and Cursor. Instinct currently has perhaps the simplest user interface of the three, but it will increasingly have to compete with companies that have enormous existing distribution advantages.

Seen on X

Other interesting stories

• Wall Street Journal: OpenAI Buys Startup Developing Smartphone Camera
• New York Times: Trump on AI and Technology (video)
• The Information: Apple Considers Server-Market Return and Nvidia Networking Technology
• SemiAnalysis: Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading
• Jasmine Li: AI Safety in China: A Primer