Welcome, humans.

Anytime “GPT-5” starts trending on X, you know OpenAI hype is at all time highs. This time around, many speculators pointed to Sam’s previous hints about the winter solstice (x.com) related to project Orion as evidence for GPT-5’s imminent arrival.

The logic goes: February’s the last winter month, OpenAI has a big Super Bowl ad, so therefore GPT-5 will launch at the Super Bowl??

Sam himself was hyping up GPT-5 in Berlin, saying he doesn’t believe he will be smarter than GPT-5 (x.com)—and he sees that as good news for humanity.

The thing is, OpenAI doesn’t actually need to release anything new tonight.

Even though Gemini 2.0 is now the top AI in the LM Arena, ChatGPT Deep Research has pretty much wowed everyone. And by everyone, we mean even famous AI skeptic Gary Marcus. Isn’t one breakthrough enough for the week?

For so long, the best quality AI couldn’t access the web. Now it can—with reasoning—and it’s pretty amazing. For example, here’s who Deep Research thinks will win the Super Bowl today. 

That said, if OA does announce something big tonight, it could be a new AI-powered device.

Here’s what you need to know about AI today:

• Big tech will spend $325B on AI this year, and we break down the implications.

• SoftBank to invest $40B in OpenAI at $260B valuation.

• Ilya Sutskever could get $20B valuation for his AI safety startup.

• DOGE proposed a new AI chatbot for gov employees.

If AI can be made for $50, what does that mean for the rest of the economy?

Big Tech is about to make it rain AI—Meta, Google, Microsoft, and Amazon collectively plan to spend $325B on AI and data center CAPEX this year.

Here’s how that breaks out: 

• Meta: $60-65B. 

• Alphabet (Google): $75B. 

• Microsoft: $80B. 

• Amazon: $100-105B—Fun fact: AWS’ cloud now makes ~$100B/year, too.

That’s a 46% jump from last year’s ~$223B splurge, btw (x.com).

All four see this spending as existential, a “once in a lifetime opportunity” to capture the market. And they’re moving in lock-step to communicate the potential value.

Meanwhile, a team of Stanford and UW researchers quietly dropped a bombshell: they recreated an OpenAI-level AI model for less than $50 in compute costs. Yeah, you read that right—fifty bucks.

Their model (called “s1”) uses a clever trick called “test-time scaling.” Instead of building a massive model from scratch, they:

1. Taught it to “think longer” about problems by adding “Wait” commands.

2. Fine-tuned it on a small dataset.

3. Ended up with something that performs similarly to OpenAI's models on math and coding tests.

The wildest part? It only took 26 minutes to train.

This isn't an isolated incident. We all remember what happened when rumors claimed DeepSeek built a top-tier model (R1) for only ~$6M.

Turns out, DeepSeek likely built its R1 model for more like $1.6B. But in an ironic twist, researchers at Berkeley were able to recreate R1’s core capabilities for only $30.

That's less than a nice dinner for two! Which reminds us: better get your V-day reservations in ASAP! We hear OpenAI has an AI for that…

Now, the catch is these models (the Berkeley model, S1 and even R1 itself) weren't built from scratch—they were “distilled” from larger frontier models. In S1’s case, it was distilled from Gemini 2.0.

Think of distillation like making espresso from already-brewed coffee: the hard work of growing, roasting, and brewing is done, but now you can make more efficient shots.

And this is where the economics of AI gets wild. After the market freaked out on R1 and dumped NVIDIA, all the tech CEOs started squawking about Jevon’s paradox. It's this funny thing that happens when something becomes more efficient and cheaper: people end up using way more of it.

For example, when cars got more fuel-efficient, we didn’t use less gas; we drove more. Or when electricity got cheaper, we didn't use less power—we invented new ways to use it. Well, Big tech is betting hundreds of billions that AI will go the same way.

According to Bloomberg Economics, there’s three ways this all will play out:

1. The Optimistic Case: Cheaper AI = more innovation everywhere. Goldman Sachs sees a 2.3% GDP boost by 2034, and McKinsey projects 5–13% growth by 2040.

2. The Pessimistic Case: AI fizzles, automating just 5% of current tasks in a decade and boosting GDP by only 1%. More like the Segway than the internet.

3. The Dystopian Case: AI works too well—but only for those who own it. Wealth concentrates, white-collar displacement follows, and widespread unrest sparks.

Interestingly, all three scenarios seem to be unfolding at once:

Pessimistically, big tech market cap surpassed 15% of US GDP (internet company market caps only hit 10%) and productivity charts show less than stellar AI gains.

Optimistically, 50% of office workers are using AI in secret, with some finding it like getting “a third of an additional person working for free.” But this hidden surge doesn’t show in official stats.

Dystopianly, eight companies have already announced AI-driven layoffs in 2025.

FROM OUR PARTNERS

Everything you need to know about AI agents

Galileo just dropped a 100-page ebook on AI agents, so you can create powerful, reliable agents like an expert:

• Match the right agentic framework for your use case.

• Evaluate and improve performance.

• Identify failure points and production issues.

Read the eBook 

Prompt Tip of the Day

Check out this person who used Deep Research as a recruiting tool (x.com).

Wanna try it yourself? Try this prompt we made, following the prompt advice from this thread—it worked pretty great for us!

Treats To Try.

1. Google AI Studio processes your massive documents, videos, and images with customizable AI settings and a huge context window… for free. Seriously, don’t sleep on this resource—we know it’s intimidating, so here’s a ~10 min tutorial.

2. ShiftApp lets you edit any text on your laptop by highlighting, double-tapping shift, and entering your prompt to transform it instantly—demo here (Mac only RN, Windows waitlist).

3. Lovable Launched is a list of the top apps built entirely with the company’s AI app builder—check them out and see what’s possible.

4. Krea transforms your sketches and images into polished artwork in real-time, and now with Krea Chat, you can edit these images through a chat interface.

5. Topaz Labs, the #1 image upscaler used by the film industry, now restores your vintage videos to crystal-clear HD with their new Starlight tool (x.com demo).

6. Silicon Valley Disposition auctions off items from tech companies (or their programs) that have gone under—it’s like the eBay of robots and bulk laptops.

See our top 51 AI Tools for Business here!

Around the Horn.

• SoftBank will invest $40B in OpenAI at a $260B pre-money valuation, paid out over the next 1-2 years, with the first payment coming this spring.

• Ilya Sutskever, OpenAI’s most famous ex-researcher, could raise a new round of funding for his Safe Superintelligence, Inc. startup at a valuation of $20B.

• DOGE (Elon’s Department of Government Efficiency) wants to develop a genAI chatbot for the US General Service Administration called GSAi to boost the productivity of its 12K employees.

FROM OUR PARTNERS

How to debug agent performance

Are your AI agents making the right decisions? Don't miss Galileo's upcoming webinar to learn how to evaluate your agents and build a continuous feedback loop to improve their performance.

Save your spot

Under the Hood

HuggingFace created an open-source Deep Research tool, but they weren’t the only ones. Here’s five others:

1. Dzhng’s version (code).

2. Nickscamara’s version (code).

3. RLanceMartin’s version—works with DeepSeek R1 (code).

4. Hwchase17’s version, GPT Researcher (code).

5. Mattschumer’s version (code).

Sunday Special: When AI gets too real…

A Cat's Commentary.