😸 AI for science is getting real.

Welcome, humans.

Doorcam footage is quite popular on Sora… the funniest kinds to us, for whatever reason, are when people experience comically horrifying weather changes:

Maybe we find these wild weather videos so relatable because we feel like we’re being swept away by an uncontrollable onslaught of heavenly force every time we try to consume a day’s worth of AI news across X, Reddit, Youtube, and Google….

Or maybe they’re just funny lol.

Here’s what happened in AI today:

1. We break down the push into physical science from the AI labs.
2. Anthropic launched Agent Skills and Microsoft 365 integration for Claude.
3. Google AI Studio combined Chat, GenMedia, and Live into a single refreshed UI.
4. DeepMind partnered with Commonwealth Fusion Systems on a fusion reactor.

Advertise in The Neuron here

Everyone’s getting REALLY into AI for science…

DEEP DIVE: Our breakdown of the latest advances in AI for the physical sciences

AI is getting close to making some major breakthroughs in both cancer research and fusion energy, and it’s making AI industry folks VERY excited about AI for science.

First, Google’s “novel” breakthrough: Google DeepMind's Cell2Sentence-Scale 27B model predicted an entirely new cancer therapy approach… then proved itself right in the lab.

Here’s what they did:

• Many tumors are “cold” (invisible to your immune system).
• Google's model screened 4K+ drugs and identified “silmitasertib” as a conditional amplifier that boosts immune signals when low-level interferon is present.
• Here’s what makes this huge: this approach had never been reported in scientific literature.
• The results = 50% increase in antigen presentation, making tumors visible to the immune system.
• Between 10-30% of the drug hits were known; the rest? Brand new discoveries.

Google has been building toward this for 10 years, since 2015 when they first applied deep learning to genome sequencing. That work produced DeepVariant, which helped complete the human genome, then AlphaMissense for predicting disease variants, and now DeepSomatic that catches cancer mutations other methods miss.

And it’s not just Google: OpenAI is also going all-in on science. They hired Alex Lupsasca, a theoretical physicist and 2024 New Horizons Breakthrough Prize recipient, as their first OpenAI for Science team member.

Here are a few science breakthroughs that OpenAI has already attributed to GPT-5:

• Sebastion Bubeck said GPT-5 Pro improved a convex optimization bound by 50% in 17 minutes (paper).
• Terence Tao used GPT-5 to solve a MathOverflow problem that would've taken “multiple hours” manually.
• Scott Aaronson published a paper where GPT-5-Thinking provided the key technical breakthrough in ~30 minutes (paper).

OpenAI also partnered with U.S. National Labs, giving 15K scientists access to frontier reasoning models for everything from disease treatment to nuclear security.

Merging AI with actual lab work. Lila Sciences just raised $115M to pioneer “scientific superintelligence” by integrating AI with physical labs, and Periodic Labs raised a $300M seed round to build “AI scientists” with autonomous powder synthesis labs.

Why this matters: Jerry Tworek, OpenAI’s VP of Research, went on the MAD Podcast and made it clear where all this is headed:

“Every week or every other week on Twitter, I do see what I think are credible reports of actual scientists using some of our reasoning models to help them perform calculations, solve hard technical problems,” he said. “Solving competitions is cool, but people solve competitions to prove they can go to actual frontier level jobs and solve new technical problems. And this is what we want from our models as well.”

Jerry Tworek, 1:05:17

What does that look like? In August, OpenAI and Retro Biosciences redesigned the Yamanaka factors (Nobel Prize-winning proteins that reprogram cells). Their AI-designed variants achieved 50x higher expression of stem cell markers and enhanced DNA damage repair, which could help us eventually slow or even reverse aging.

And if you think solving aging is ambitious, DeepMind just announced a partnership with Commonwealth Fusion Systems on SPARC, which aims to be the first fusion device to achieve net energy (breakeven). Their TORAX simulator runs millions of virtual experiments, and RL agents are learning to control plasma at 100M+ degrees Celsius. If this works, it could lead to the ability to control infinite energy. Who knows, that might finally be enough to power all of OpenAI’s datacenters! Read more here.

FROM OUR PARTNERS

Webinar - Forward to 2026 — 3 AI trends that won't take you backwards

AI has become the loudest topic in customer experience—but many CX leaders are realizing that pilots stall, tools pile up, and “faster” doesn’t always mean better.

In this session, Liz Miller (VP & Principal Analyst, Constellation Research) and Gladly’s CX experts will unpack the key AI shifts shaping 2026—like why smarter decisions are starting to matter more than speed—while separating hype from what will truly drive impact.

The goal: give CX leaders clarity on which trends will drive meaningful outcomes—and which ones risk pushing teams backward.

Call to action text: Register now

Prompt AI Skill Tip of the Day

Anthropic just launched a new resource called Agent Skills and new Microsoft 365 integration, two features that make Claude way more useful for actual work.

Skills are basically instruction folders Claude loads only when needed. Go to Settings > Capabilities > Skills, enable “skill-creator,” and Claude guides you through building custom skills interactively (docs for how). Skills can stack (Excel + presentations simultaneously) and include executable code for reliable outputs.

Available for Pro, Max, Team, and Enterprise. Browse example skills here.

Now, for Team and Enterprise customers, Claude now searches SharePoint, OneDrive, Outlook, and Teams through one MCP connector. The Enterprise search feature creates a shared org-wide project that pulls from all connected tools; ask about company policies and get answers synthesized from emails, docs, and chat threads.

Our favorite insight: Skills teach Claude how to work (your processes, brand guidelines, workflows), while M365 integration gives it what to work with (your actual company data). Together, you get a specialist who's already onboarded and has access to everything. No more “let me find that document” or “I need to check Teams first”, just ask Claude. FYI, admins must enable both features before teams can use them.

We wrote about this a bit more in depth here!

Treats to Try

1. Google AI Studio got a refresh (above) that combines Chat, GenMedia, and Live into a single UI so its easier to work with them all.
2. Dedalus helps you build complex AI agents across any model and tool in 5 lines of code, handling model handoffs, tool chaining, and deployment so you don't have to write hundreds of lines or battle with infrastructure (raised $11M).
3. Liberate automates insurance back-office operations with AI, helping clients achieve 15% sales increases and 23% cost reductions (raised $50M).
4. General Intuition develops AI models with spatial-temporal reasoning using Medal's dataset of 2 billion video game clips (raised $134M).
5. KAYAK AI Mode lets you find and book travel by typing questions like "Find me a family-friendly hotel in Rome with free breakfast" instead of using search filters (read more).
6. Strella offers interview technology for customer research, with companies using the platform reporting 85% better sales growth than peers (raised $14M).

Around the Horn

• Apple lost Ke Yang, head of web search efforts, to Meta in a continuing trend of high-profile AI departures.
• Europe deployed AI-driven mega water projects in its driest regions, raising sustainability concerns among environmental experts.
• Spotify partnered with major labels to develop artist-first AI music products with fair compensation and opt-in controls.
• Pinterest added feed controls to reduce AI-generated content after discovering up to 75% of wedding search results were AI-generated.
• California passed SB-53 requiring frontier AI developers to publish risk frameworks and report safety incidents within 15 days—check out this and more safety-related AI insights from the AI Safety newsletter!

FROM OUR PARTNERS

Stuck explaining what you want to an AI image generator?

Recraft just launched Chat mode—think ChatGPT meets a Photoshop canvas. You can describe changes in plain language, and Recraft updates your image instantly.

It understands your project, remembers past edits, and suggests smarter refinements as you go. From re-coloring and background edits to vector tweaks, Chat mode makes iteration feel like collaboration. Used by Netflix, HubSpot, and Asana to create production-ready designs at speed.

Check it out.

Intelligent Insights

• This is a great convo between Sinead Bovell and Kate Crawford on the debate around AI and energy and the future is defined by the choices we make today.
• Chinese AI models overtook Western leadership in 2025's open-source landscape while synthetic data slashed post-training costs by 99%.
• An economic analysis of the AI bubble reveals trillion-dollar investments generating only $45B annually, with GPUs burning out faster than claimed depreciation schedules.
• This comprehensive guide on solving the AI memory crisis reveals how intelligence has outpaced memory capabilities by 600x, creating a critical bottleneck spawning an entire industry.

A Cat’s Commentary