Welcome, humans.

ICYMI: We just sat down with Kari Briski, Vice President of Generative AI Software for Enterprise at NVIDIA, who revealed something wild about NVIDIA's Nemotron model family.

In our conversation, we learned why NVIDIA is giving away their entire AI playbook: 500+ models, training data, algorithms, even the “gym environments” they use to make AI smarter—while competitors guard their secrets like Fort Knox.

She also walked us through exactly what hardware you need to run ChatGPT-level AI on your laptop (spoiler: 18GB GPU for their smallest model) and why enterprises are ditching cloud AI to build their own “deep researchers” that never leave their servers.

If you've been waiting for truly private, customizable AI that you control completely, this interview is for you! Watch / Listen now on: YouTube | Spotify | Apple Podcasts

Here’s what happened in AI today:

1. Andrej Karpathy predicted progress towards AGI will take another decade.

2. Meta secured $27 billion in financing for its new Hyperion datacenter.

3. Wikipedia reported 8% traffic decline due to AI summaries and social media.

4. NVIDIA and TSMC announced the first completed U.S.-made wafer for AI chips.

Andrej Karpathy says AGI is “ten years away”… here’s why.

DEEP DIVE: Our breakdown of all of Karpathy’s key quotes and predictions from this epic 2 and a half hour talk.

In a wide-ranging interview on the Dwarkesh Patel podcast, AI legend Andrej Karpathy threw some cold water on the entire industry: forget the hype about the “year of agents”… we should be thinking about the “decade of agents.” His reasoning? Today's AI has serious cognitive deficits that won't be fixed anytime soon.

Keep in mind: this is coming from the guy who coined the term “vibe-coding.”

Karpathy, who previously led AI at Tesla and was a founding member of OpenAI, argued that large language models like ChatGPT are nowhere near being the reliable, intern-like assistants we imagine.

Here's his brutal assessment:

• Reinforcement learning is “terrible” and “stupid” because it rewards lucky guesses the same way it rewards actual reasoning.

• Coding agents aren't useful yet; they get confused by custom code and bloat projects with defensive boilerplate.

• Perfect memorization is holding models back; unlike humans, who generalize because we forget, LLMs are distracted by their encyclopedic recall of the internet

• Getting from 99% to 99.9% reliability takes as long as getting from 0% to 90%—and we're nowhere near 99% yet.

Naturally, Gary Marcus took a victory lap.

Karpathy's take challenges the $14 trillion that's been added to tech valuations since ChatGPT launched. The market has priced in a world where AI automates all digital work. But if Karpathy's right, that's a decade-long grind, not a two-year sprint.

WHY IT'S IMPORTANT: This leads us to believe we need to reframe the “artificial intelligence” discussions into “augmented intelligence” for at least the next ten years. Instead of thinking about robots doing everything for us, think about what happens when we use AI to accelerate our own thinking, our own actions, and our own science.

As we wrote on Friday, scientists can already do incredible things with the augmented intelligence gains from today's AI models. These gains will compound over a decade through better scaffolding (software that directs the AI to be more useful). So we'll still get incredibly valuable leverage from augmented intelligence.

But it won't be “AGI” like it’s been sold to us so far; according to Karpathy, there’s nothing “general” or “intelligent” about today’s language models, really. So reset your expectations: no AGI until 2035.

Instead, think of today’s AI as compressed human knowledge you can tap and remix to your needs. Like a mini Wikipedia in your pocket. But would you trust Wikipedia to run your business for you? If not, you probably don’t want todays’ AI models to either.

For more insights, watch the full interview, read the full transcript on Dwarkesh’s blog or check out our complete deep dive breaking down all of Karpathy's key quotes and predictions from this epic 2.5-hour talk.

FROM OUR PARTNERS

What the Fastest Growing SaaS Companies Use for Enterprise Identity

Enterprise customers demand SSO, directory sync, granular permissions, and audit logs built to strict compliance standards. Building them in-house takes months and pulls focus from your core product.

That’s why companies like OpenAI, Cursor, and Vercel use WorkOS, the developer platform powering auth, identity, and security for B2B apps.

- Modular, developer-friendly APIs for SSO, SCIM, RBAC, and more

- A hosted Admin Portal for seamless customer onboarding

- Compliance features like SOC 2 and GDPR built in

1000+ companies already rely on WorkOS. Your first million MAUs are free.

See why the best build on WorkOS →

Prompt Tip of the Day

Here’s a piece of prompt advice nestled a few hours into the Karpathy interview: Use ChatGPT as your “dumb question assistant.””

When reading a biology paper recently, Karpathy uploaded it to ChatGPT and asked all his basic questions about it in the context window. After working through the confusion, he shared the entire thread with the paper's author.

Why? Seeing beginner questions helps experts explain better. Karpathy said he'd love if people shared their “dumb ChatGPT conversations” about his tutorials—it shows him exactly where learners get stuck.

The takeaway = don't be embarrassed by basic questions. Ask ChatGPT first with the material in context, then share that conversation with the expert (or a coworker/boss, if the way they communicate or the work they create confuses you). You'll learn faster AND help them become a better teacher / communicator. But uh, make sure they’re open to feedback first!

Treats to Try.

1. Luzia is your AI study buddy that helps you solve math problems, create images, get homework help, and handle life questions—completely free on mobile

2. Lazy captures anything to Notion with one shortcut (articles, tweets, videos, emails) without leaving your current app.

3. Snap runs usability tests in 5 minutes by simulating AI users who test your designs and give you actionable feedback.

4. Emergent turns your English description into a working full-stack app using AI agents that plan, code, test, and deploy.

5. KaneAI writes and runs software tests in plain English across web, mobile, and APIs—no coding needed.

6. Waydev AI answers questions about your engineering metrics by analyzing GitHub, Jira, and other dev tools in plain language.

7. Endless summer fakes your vacation pics for you so you can look like you went on an awesome vacation to Ibiza when instead you were locked in working.

Around the Horn

• NVIDIA and TSMC announced the first “completed U.S made wafer”at TSMC's Phoenix, Arizona facility, which will eventually become Blackwell AI chips.

• Meta secure financing for its new Hyperion 2.2 gigawatt datacenter, Hyperion, to the tune of ~$27B from a private equity firm called Blue Owl Capital; separately, Amazon revealed its own 960 megawatt datacenter called Cascade Advanced Energy Facility, set to come online “sometime in the 2030s.”

• Wikipedia said traffic to its website declined ~8% year-over-year due to genAI summaries and young people using “social media” (YouTube and TikTok) as search engines to seek information.

• Japan’s government made a formal request for OpenAI to stop “copyright infringement” on Japanese anime, calling anime “irreplaceable treasures”… however, Japan has very liberal copyright laws for training AI for commercial use, aiming to be “the most AI-friendly country.”; perhaps the rules need to specify the differences between using protected IP via inference vs training?

• WhatsApp updated its terms to ban “general-purpose” AI bots starting January 15, 2026, but will still allow AI that serves customers, as opposed to services who just use WhatsApp as distribution for a chatbot (details).

• The Hollywood Reporter published a great timeline of the launch of Sora 2, with reactions from agencies like WME leading up to and after the launch.

• While TechCrunch said OpenAI needs to generate $1T in value over the next five years, Ed Zitron says OpenAI actually needs $400B in the next 12 months to actually complete any of these deals; this is why Karpathy’s message is so urgent… we need to get REALISTIC about what AI can and can’t do, and act accordingly.

Sunday Special

Anil Dash has the best take on the AI industry atm, which rhymes with a lot of what Karpathy said: the majority view inside the industry is that large language models have utility, but are being “massively overhyped” and “forced on everyone” without the ability to “focus on legitimate uses where they might add value.”

It’s a fantastic piece, and it sums up and expands on our own feelings quite well; we like talking about the new things AI can do, and teaching you how you can use these new tools, but we are also trying to actively present a realistic and balanced view of what happens next. That’s why we joke about the AI bubble (and also take it seriously), and highlight the industry’s moves with a healthy dose of skepticism while still sharing perspective pieces from skeptics and maxis alike.

Ultimately, everyone must make their own decision about where to implement this technology, and how it can or can’t help them in their individual journey. But there is a real and present danger of over-indexing on today’s technology from an investment standpoint, which as Anil writes, limits “the universe of possibilities for the future.”

What that means (to us anyway) is that there are an unlimited number of ways technological progress can happen, but if we force all of our money, people, and attention into any one path or company, we risk losing out on the infinite other possible futures and outcomes we could pursue.

A Cat’s Commentary