🧙🏼 OpenAI's o3 mini is here

Howdy, wizards.

What a week. The Chinese company DeepSeek made huge waves in the AI industry and beyond. OpenAI dropped o3-mini, and the supercharged o3-mini-high. Sam Altman spilled some juicy tea about upcoming releases in a Reddit AMA.

Let’s get into it, shall we? Here’s what’s brewing in AI.

DARIO’S PICKS

1. OpenAI drops o3-mini & o3-mini-high

OpenAI first announced their new reasoning model o3 in December. The little brother of the model family—o3 mini—is now officially here, and it’s quickly becoming a favourite among ChatGPT users.

o3 mini is the smaller, faster and cheaper version of the upcoming (full) o3 model. It’s optimized for STEM domains, and also top of the line when it comes to coding.

o3 mini is released both in ChatGPT and in the API (so devs can start building apps with it) right away.

o3 mini in ChatGPT:

• Available now to Plus/Team/Enterprise/Pro users, there’s two new options in the model picker: o3-mini and o3-mini-high. The difference is in how hard the models thinks before answering, and how many messages you can send:

  • o3-mini: reasoning level set to “medium”, limit of 150 messages per day

  • o3-mini-high: reasoning level set to “high”, limit of 50 messages per week

  • ChatGPT Pro users (the $200/month tier) have unlimited messages to the new model

• Free users get free 10 messages to o3-mini per day by clicking the “Reason” button

o3 mini in the API:

• o3 mini has the cost efficiency and latency of o1 mini, but is a more advanced reasoning model.

• It supports some features for developers that makes it way easier to build production-ready apps with, including function calling (can use 3rd party APIs), structured outputs and developer messages.

• When using the API, you can set the aforementioned reasoning level yourself, to either “low”, “medium” or “high”—whichever suits your use case (e.g. you can choose to optimise for speed when problems aren’t that complex).

‎ Why it matters‎ ‎ I tested o3-mini-high on some fairly complex situations where it had to calculate my taxes based on different scenarios, and it rocked! GPT-4o spewed out nothing but faulty calculations (which looks convincing on the surface but are not correct) on the same questions.

However, since o3 mini doesn’t yet support file uploads, it can’t reason over documents and images. Pretty sure I’ll find myself still using o1 a lot until it does. It’s coming soon, though.

Oh, and if you’re wondering when the full o3 model is expected to launch, I interpret OpenAI’s Reddit answer to mean ≈ 1-3 months from now.

TOGETHER WITH GAMMA

An entirely new way to present ideas

Gamma’s AI creates beautiful presentations, websites, and more. No design or coding skills required. Try it free today.

DARIO’S PICKS

2. DeepSeek shakes everything up

DeepSeek, a Chinese company founded by a hedge fund manager with a team apparently building AI models as a side project, trained a really advanced, open weights AI model. They did so with dramatically less compute and in less time compared to the leading, US-based tech companies — with an order-of-magnitude reduction in costs. DeepSeek even reached the #1 most downloaded app on Apple’s app store this week.

This led to a huge selloff in US tech stocks with Nvidia shaving off $600b in a market cap in a single day and even Trump calling it a wake-up call for the industry.

The impact of DeepSeek’s release is multifaceted and extends far beyond mere technological advancement—into topics like economics, politics, privacy, censorship, export controls, open source, and more. Or as The Dude would say: lotta ins, lotta outs, lotta what-have-yous.

Here’s the main deets:

• Performance-wise, DeepSeek’s model, R1, is a very capable reasoning model comparable to OpenAI’s o1 and o3. While it doesn’t outshine the top models on the market including o1/o3 and Claude 3.5 Sonnet, it stands as one of the best free AI models. I would have said the best free model if it weren’t for OpenAI making o3-mini available to free users this week (albeit with very limited usage).

• Arguably the most notable thing about DeepSeek’s R1 is that they managed to train it super cheap, based on reinforcement learning and efficient architectures like Mixture-of-Experts. The model was reportedly trained for about $5.5M—vastly lower than GPT-4’s estimated $100M. However, notable voices such as Anthropic CEO Dario Amodei suggest that while the actual cost is probably lower than for US-based top models, the actual figure isn’t even close to what DeepSeek claims. Amodei also says the cost reduction isn’t that surprising but rather a “expected point on an ongoing cost reduction curve”.

• Something ironic in all this is that the US has been restricting exports of its most advanced AI chips to slow China’s progress—a move that now appears ironic as DeepSeek’s breakthrough rattles the market (a sentiment echoed on Reddit 😄).

• Privacy and free speech remains concerns with DeepSeek. The model is completely tight-lipped when discussing the Chinese government, and a recent data breach apparently exposed 1M log files. For increased security, it’s totally possible to run it locally on your laptop through a tool like LM Studio, but the performance will be very much reduced.

  • Related: DeepSeek’s virality this week might’ve prompted OpenAI to release ChatGPT Gov — a specialized ChatGPT version for US agencies that contains their data in a secure Azure environment. According to the company, 3,500 US Government agencies are already using AI (here’s a list I’ve with 21 of them).

• OpenAI is saying DeepSeek copied them, using their model to train their own. True or not — it’s meme-worthy:

‎ Why it matters‎ ‎ DeepSeek’s model shows that there’s no wall in sight yet when it comes to acceleration of AI. It shows it’s possible to train intelligent models with a fraction of the resources we thought we needed. It increases the competition at the model layer of the industry, and consumers are left with free, powerful AI models. It also seems to have made OpenAI rethink it’s direction of closed models.

Short-term this is bad news for NVIDIA’s stock price, no doubt. The company sits at the heart of the industry with its advanced AI chips that everyone relies on to train their models. And everyone is now scared they’ll not have the same demand since you need less compute for the same result. Another way to look at this is that A) it expands the market as a lot more companies will now have the money to train models from scratch and achieve great outputs, such as highly advanced models for very specific domains and B) the big companies will simply use the improved economics to train even more powerful models — not cutting costs. Not financial advice, btw.

TOGETHER WITH HUBSPOT

3. HubSpot's Starter Bundle for Startups

Unlock the potential of your growing startup with HubSpot's Starter Customer Platform. For a discounted price of just $20/month, unlock access to the Starter edition of HubSpot’s six core products for marketing, sales, and customer service – powered by HubSpot’s Smart CRM – all for the price of one. Built for startups like yours, HubSpot Starter has all the essential tools you need to scale.

Try HubSpot

DARIO’S PICKS

4. OpenAI’s Reddit AMA

Six key people from OpenAI did a Reddit AMA (Ask Me Anything) session yesterday and, similar to last time they did it, teased some pretty cool things.

Check out the full AMA to see all the answers.

Here’s the answers I found most interesting (ranked order):

• OpenAI is thinking of going more, well, open. A Redditor asked if they’re going to release model weights and more research, to which Sam Altman said they’re discussing it, have been “on the wrong side of history” with their current approach, and need to figure out a different open source strategy. He also noted not everyone internally agrees on this and that it’s not currently their highest priority.

• Very soon, OpenAI’s models will start showing much more detail of how they are thinking soon, inspired by DeepSeek’s R1. According to CPO Kevin Weil, though, it’s a balance because showing all of the thinking makes it easier for competitors to copy.

• Operator on the regular Plus plan is currently months away. They’re currently held back by costs, and are currently training specialised Operator models that are faster and cheaper.

• It might become easier for EU companies to build with OpenAI. They’re currently working on a EU data residency offering, and will be doing a broad rollout of this at some point.

• OpenAI is working on an image generator that uses GPT-4o. To be released in a couple of months, and claimed to be “worth the wait”.

• Sam Altman thinks that a recursive self-improvement for AI looking more like a hard take off rather than gradual improvement is more likely than the thought a couple of years ago (translation: AI might get much better — very fast).

• OpenAI has a lot of components for a complete AI automation suite already: voice, video, screen share, tasks and operator. According to Sam Altman, the most critical missing pieces for a complete AI experience are improved reasoning capabilities, automation in any environment (not just the browser), and the ability to understand and use any apps. Also says, “continuous video in and video out would be really cool”.

• Altman thinks the aspect of AGI that will improve quality of life the most is the ability to accelerate the rate of scientific discovery.

• There will be a new model at some point named GPT-5, not GPT-5o. Altman has highlighted that solving their naming problem is a key goal for 2025.

QUICKFIRE HEADLINES

Side notes

Here are the other essential updates on AI you shouldn’t miss this week, that I didn’t have time to cover in-depth.

• OpenAI is partnering with US National Laboratories giving scientists access to its most advanced AI models for research, including nuclear weapon security.

  • Related: The advocacy group Bulletin of the Atomic Scientists just moved their “Doomsday clock” to 89 seconds to midnight. It captures threats including climate change, nuclear weapons, instability in the Middle East, pandemic threats, and use of AI in military operations.

• Anthropic CEO Dario Amodei weighs in on DeepSeek in a new essay

• New, clearer guidelines on AI-generated works from the US copyright office

THAT’S ALL FOLKS!

Affiliate disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes link to products and other newsletters. Please assume these are affiliate links. If you choose to subscribe to a newsletter or buy a product through any of my links then THANK YOU – it will make it possible for me to continue to do this.