🤓 This Week in AI

PLUS: AI Deepfakes Are Now Too Convincing for Comfort

OpenAI is eating the world.

It dropped two major launches, announced new partnerships—and that was just by Tuesday.

This was me all week just trying to keep up: 🫠 🫠 🫠

But the thing that really stopped me in my tracks didn’t come from OpenAI.

It was from TikTok’s parent company, ByteDance, and it’s both remarkable and deeply unsettling.

They showed off their new AI system, OmniHuman, which can create the most realistic human avatars and deepfake videos I’ve ever seen.

Beyond generating convincing fake images of real people, it can also create high quality animated characters.

All you need is an image and an audio or video clip. The model can even edit existing videos—adjusting facial expressions, gestures and even limb movements.

I know the shared samples are cherry-picked, but still—WOW.

Here is a fictional Taylor Swift performance:

And an animation example:

There was also a clip of Swift singing in Chinese, and another of a TED Talk that never happened, but they’ve since been taken down (likely for copyright reasons.)

You can check out the rest of the demos here. They’re truly astonishing.

ByteDance hasn’t released the system yet, but these demos have officially erased the line between real and AI-generated videos.

As similar powerful systems inevitably become available, the real danger isn’t that deepfakes will fool us—it’s that they’ll make us doubt everything, even the truth.

Okay, now that we’ve all had an existential crisis, let’s get into the rest of this week’s AI updates.

Here’s what you need to know about AI this week (clickable links appear in orange in emails and underlined in the Substack app):

---

OpenAI had two big releases this week that are worth your attention, because they give us a glimpse at how AI is reshaping the very definition of ‘expertise.’

One of them left me completely stunned when I tried it—seriously, I haven’t felt this way about an AI tool since ChatGPT first launched.

I’ll get to that soon, but first, let’s break down both launches.

1️⃣ o3-Mini: OpenAI’s New, Smarter and More Reliable Reasoning Model

Unlike most large language models, reasoning models take time to “think,” using a step-by-step technique called chain of thought to break down problems logically. This means Reasoners are capable of solving much harder problems.

Responses take longer, but they’re higher quality, more reliable and less prone to mistakes, and their capabilities are quickly expanding.

Also, the longer they “think,” the better their answers get (though the rate of improvement slows as they think longer).

o3-Mini:

• Breaks down complex tasks into smaller manageable parts

• Checks its own work to spot and correct mistakes before responding

• Adapts and tries different approaches if the ones it tries first don’t work

It’s ideal for math, coding, sciences and complex problem solving like analysis, strategy development and decision-making.

It makes 39% fewer “major mistakes” on “tough real-world questions.”

And even scored 93% on OpenAI’s research engineering interview questions (basically, it’s acing job interviews at OpenAI) 👀

I’ve been testing o3-mini out all week, and its performance is in line with o1—which was already very strong for tackling complex problems—but with noticeably more consistency.

How to use o3-mini in ChatGPT

If You’re a Plus User ($20/month):

💬 150 messages per day (which is a LOT with a model this powerful)

👉 Just select “o3-mini-high” from the model picker at the top of the ChatGPT interface.

—

If You’re a Free User:

💬 Limited access (~10 messages per day).

👉 Click the Reason button under the message composer. You can also combine the search and reasoning capabilities by choosing both buttons (see 👇).

2️⃣ Deep Research: A New AI-powered Agent for Research

Deep Research brings expert-level research capabilities to ChatGPT.

It’s an AI agent designed to handle complex, multi-step research and analysis projects independently—turning what used to take hours (or even weeks) into minutes.

Here’s how it works: You give it a prompt, and it finds, analyzes, and combines data and insights from various online sources into an in-depth and detailed report on any topic.

Powered by OpenAI’s most advanced (yet unreleased) reasoning model (o3), it:

• Asks clarifying questions to better understand your needs

• Searches across numerous and diverse sources

• Adapts its research strategy based on what it learns in real-time

• Compiles and analyzes findings into comprehensive reports

• Cites all its sources for easy reference and verification

In the meantime, you can step away or work on other tasks—you’ll get a notification once the research is complete.

Or you can watch it do research in real time, following its thought process as it digs around through various sources.

Searches can take from about 5 to 30 minutes, depending on the complexity of the topic, and the reports can run up to 16,000 words.

You can also ask it for specific formats: executive summaries, comparison tables, bullet points, strategic frameworks, phased roadmaps, and more.

It’s particularly good at finding niche information that would involve multiple steps across numerous websites and can analyze text, images, and PDFs.

You can also upload your own files—like marketing strategies, roadmaps, data sets, or internal documents—and Deep Research will analyze them alongside external sources to deliver more tailored insights.

In the next few weeks OpenAI will also be adding embedded images, data visualizations and other analytic inputs for more clarity and context.

🥇 What It’s Good At

• Finding hyper-specific information that is un-Googleable like tracking down that one poem you vaguely remember from years ago

• Market Research

• Trend and competitive analysis

• Financial, legal and academic research

• Research ahead of expensive or complex product purchases

💡 To get the best results, use your prompting knowledge and best practices.

For the many of you who’ve attended my workshops, you’ve seen firsthand how a strategic approach to prompting can completely transform the quality of your results—often in ways that surprise even experienced power users.

Provide as much context and specifics about your goal as possible. Give it rules and constraints like “only use sources from the past two years” or “focus on U.S.-based companies only”.

Basically, treat your prompts like an agency brief—or a project brief. A really good one, with all the context, background, and clear expectations.

Oh, and yes—I know I promised to announce another workshop on a Friday afternoon. I haven’t forgotten! I know many of you have been asking (and waiting). I’m just trying to lock in a date that works for those of you who reached out directly.

If you’re interested, shoot me a note so I can make sure I’m picking a date that works for as many of you who really want to attend. I’ll share the details soon (but you can find general info here.)

⚠️ Limitations to Keep in Mind

• Occasional hallucinations or mistakes (though at a much lower rate that other ChatGPT models)

• Struggles with distinguishing credible information from rumors

• No access to paywalled articles (but might find workarounds)

• Blocked from YouTube, Google, and The New York Times

OpenAI expects all the first two issues to improve quickly with more usage and time. But for now, you’ll need to review, fact-check, and verify its output.

Luckily, reviewing takes way less time than doing the research yourself—and with all the sources linked, double-checking is quick and easy.

Access is currently limited to Pro users only ($200/month) but will soon expand to Plus users ($20/month) soon.

I highly recommend checking out this 20-minute live demo to see and fully understand its capabilities. You can also find more specifics on OpenAI’s Deep Research FAQ page.

🤯 What Happened When I Tried It

I had the chance to test Deep Research with my friend Parth, who is on the Pro tier ($200/month)—and I was blown away.

We asked it to help us position an AI app he recently built and develop a monetization strategy.

Not only did it conduct extensive research, but it also delivered a comprehensive strategy that felt like weeks of work from a top consulting team.

It redefined the target audience with detailed sub-segments, defined key differentiators, and crafted a positioning strategy.

From there, it laid out tiered subscription models, growth opportunities, partnership suggestions, revenue projections, a phased roadmap, and even surprisingly strong marketing copy.

OpenAI positions Deep Research as a research agent, but that’s underselling it—it’s more like having a team of experienced strategists, analysts, and consultants all working in sync.

I AM IN AWE. This definitely feels like another ChatGPT moment.

And it’s not just me. Others (especially senior execs) are posting about similarly jaw-dropping experiences—like this one from Runway CEO Siqi Chen, who said Deep Research combined with o3-mini outperformed the $150k private research team he hired to study his daughter’s rare medical condition. 👇

I can’t wait for the launch of this agent for Plus users ($20/month) next month, which will give us 10 queries a month. Trust me—you’ll want to try it for yourself.

🤔 What Does All of This Mean for Us?

These two launches mark a shift: we now have AI systems capable of performing work that once required teams of highly-paid experts, analysts, and consultants.

But AI’s ability to produce this level of work doesn’t eliminate the need for experts.

It shifts their role from doing the work to leading it—setting the direction, defining the problems, asking sharper questions, spotting gaps, and making the judgment calls AI can’t.

The value of human work isn’t disappearing. It’s moving upstream.

---

OpenAI makes ChatGPT Search free for everyone—No account required.

The move positions ChatGPT as a direct competitor to traditional search engines like Google and Microsoft Bing.

It also marks the final stage in ChatGPT search's rollout, which began with its launch to paid subscribers in October 2024, expanding to free accounts in December.

If you missed it, I did a deep dive on how ChatGPT Search works and what sets it apart from traditional search engines in this edition.

If you’re in marketing or PR, this is a must-read—AI search is transforming how brands get discovered, capture attention, and engage with audiences online.

---

OpenAI and Japanese tech giant SoftBank are forming a new company called SB OpenAI Japan.

The deal: This is a 50-50 partnership with SoftBank committing to pay as much as $3B a year to incorporate OpenAI’s latest models into its own enterprise products.

OpenAI has also teamed up with Korea’s Kakao to develop a Korean-language assistant and integrate ChatGPT into KakaoTalk, South Korea’s dominant messaging app with over 40 million users. Beyond expanding OpenAI’s reach in Asia, this move gives OpenAI access to Korean-language data to help train and improve its models.

---

.OpenAI has partnered with the California State University system (CSU) to bring ChatGPT to more than 500,000 students and faculty in “the largest deployment of ChatGPT to date.”

---

With AI-written books on the rise, the Authors Guild is launching a “Human Authored” certification to verify books written by people. Certified books will be listed and searchable in a public database, with the program starting for Guild members and expanding over time.

But what exactly counts as “human-authored” these days?

Some of the most brilliant writers I know are using AI—not as a ghostwriter, but as a creative partner: to brainstorm, get feedback, test new angles, sharpen dialogue.

Does that make the work less human? Or serve as a reminder that tools don’t make the art—the artist does.

If AI could write great books, we’d all be bestselling authors by now.

---

🤔 Should you be allowed to use AI to apply for a job—at an AI company?

Anthropic, the AI company behind Claude, asks candidates to confirm they haven’t used AI in their job applications.

The application puts it like this:

“While we encourage people to use AI systems during their role to help them work faster and more effectively, please do not use AI assistants during the application process. We want to understand your personal interest in Anthropic without mediation through an AI system, and we also want to evaluate your non-AI-assisted communication skills. Please indicate 'Yes' if you have read and agree.”

Basically, Claude can help you do the job—just not get the job.

---

Adobe’s Chief Strategy Officer joins top independent film company A24 as a partner, leading technology and innovation projects. His focus will be on providing tools and resources (or AI 😁) to help filmmakers bring their visions to life.

---

Can a character feel real without a human behind the voice?

Hank Azaria spent decades giving life to The Simpsons characters. He argues while AI might replicate his voice, it can’t replicate the soul, heart, and vulnerability behind it.

---

In case you missed last week’s edition, you can find it 👇:

That's all for this week.

I’ll see you next Friday. Thoughts, feedback and questions are always welcome and much appreciated. Shoot me a note at avi@joinsavvyavi.com.

Stay curious,

Avi

💙💙💙 P.S. A huge thank you to my paid subscribers and those of you who share this newsletter with curious friends and coworkers. It takes me about 15+ hours each week to research, curate, simplify the complex, and write this newsletter. So, your support means the world to me, as it helps me make this process sustainable (almost 😄).