Claude cracks 87-year-old math problem

Welcome back, Superhuman. The singularity is inching closer. In May, ChatGPT helped disprove an 80-year-old Erdős conjecture. Just a month later, Claude has answered with a major math win of its own, disproving a problem that’s stumped human researchers for nearly a century.

Today: Google is embedding Gemini in chips, how to build and publish websites with ChatGPT Sites, and get the latest prompts and trending social posts.

1. Claude Fable 5 helps crack a math problem open since 1939: Anthropic researcher Levent Alpöge — a Harvard-trained mathematician — posted a counterexample to the Jacobian Conjecture with help from Claude Fable 5. Alpöge’s result was verified by several prominent mathematicians, including Stanford’s Jared Duker Lichtman. The milestone highlights AI’s growing prowess in math, especially when wielded by top minds. Read more.

2. Google is reportedly embedding Gemini right into its next-gen chips: Alphabet is working on "Frozen v2," a new chip that bakes parts of its Gemini model straight into the silicon. The new design would reduce the amount of data required to answer queries, potentially making it 6X to 10X more efficient at serving AI responses than the company's current TPUs. The new chip is aimed at easing a major internal compute shortage, although it’s not expected to ship until 2028.

3. Washington's AI safety office gets another leadership change: Dr. Chris Fall has stepped down as director of the Center for AI Standards and Innovation, the federal agency that stress-tests AI models for risks like cyberattacks. It’s the third person to lead the office in three months: Collin Burns lasted four days, Fall served for three months, and now National Institute of Standards and Technology Director Dr. Arvind Raman is stepping in as acting director.

Slack powers over 3 million workflows everyday. And with Slack’s new Workflow Builder, teams can build unique automations and increase productivity by ~35%.

No coding required: About 80% of Slack Workflow Builders are non-technical users, meaning you can automate your sales, marketing, customer service, IT, and more with just clicks. 

The results?

And new AI tools, keyword triggers, and Salesforce integrations makes automating easier than ever before.

See the guide 30 Ways to Automate Work in Slack here.

China closes the frontier gap. Two Chinese labs just unveiled frontier-level models at a fraction of traditional costs: Moonshot AI’s Kimi K3 and Alibaba’s Qwen 3.8. In theory, it’s exactly the kind of release that should bring cheaper intelligence to everyone and help ease AI's compute crunch. But in practice, the scenario is playing out much differently.

Moonshot hits its compute limit. AI continues to experience a Jevons Paradox scenario, where people use models more frequently as they get cheaper. Moonshot AI just experienced this firsthand. Within 48 hours of launching, the lab was forced to pause new subscriptions after demand blew past its compute capacity. In theory, any organization can download Kimi K3’s weights once they become available and run the model locally. But there’s a major catch.

Kimi K3 requires a specialized rig. You can’t run Kimi K3 on a laptop. The lab recommends using a "supernode” of 64 or more datacenter accelerators to run the model. And even if a company is ready to write the check, the chips aren't there to buy: ordering your own data-center GPUs today means waiting roughly 36 to 52 weeks for delivery, thanks to an industry-wide backlog.

Compute is still the industry’s bottleneck. For now, AI’s limiting factor isn’t how capable the models are. It's getting the chips to run them. Fortunately, many labs are already working on one potential solution: smaller models that can run on hardware companies already own.

Sample Prompt: @Sites Create a website to build an interactive project tracker for my marketing team. Include task owners, deadlines, priorities, progress, and filters for status.

Sample follow-up prompt: Add a dashboard showing completed, overdue, and upcoming tasks. Make the design cleaner and more professional.

💡 Tip: Return to the same project whenever you need to update the Site. Describe the changes you want, preview them, and publish the updated version.

Block AI-powered phishing and ransomware, manage team apps, and prevent sensitive data from leaking through AI tools—all from one secure workspace built by Palo Alto Networks.

Start Free Trial

✏ De-AI Your Writing: Want your LLM to sound more human? Try adding these 6 writing rules to your favorite chatbot’s instructions. They force it to use active voice, clear sentences, and avoid metaphors (10K bookmarks).

👀 Kimi in Action: Explore some of the truly insane projects people have already created using Kimi K3. Our favorites? An exploration-based video game, an Animal Crossing clone, and a 3D globe dashboard (1.5M views).

🦾 Best Models: This list breaks down which models are best for completing different tasks. It’s pulled in 4M views and 3K bookmarks, so the internet seems to agree with its choices.

📚 Bookmark Worthy: This Claude Code skill racked up 39K bookmarks with a simple promise — it cuts the fluff out of AI’s answers. Try the skill here.

🤖 Ramp Router: Ramp cut its LLM costs by 30% with an internal tool that routes each request to the best model for the job. Cheap models for simple tasks, frontier models for hard ones. Now it’s opening the tool to the public.

^{* indicates a promoted tool, if any}

Lucy 2.5 is a new model that edits live video streams with barely any lag, letting you conjure up special effects on the fly. Watch the viral launch video here (4.5M views).

Grow customers & revenue: Join companies like Amazon, HubSpot, and Salesforce. Showcase your product to our 1M+ readers and 2M+ followers on socials. Get in touch.

Until next time — Zain, Theodore, & the Superhuman AI team.