- AI Valley
- Posts
- The End of Search, The Beginning of OpenAI’s Deep Research
The End of Search, The Beginning of OpenAI’s Deep Research
PLUS: OpenAI has unveiled o3-mini
Together with
Howdy! It’s Barsee again.
Happy Monday, AI family, and welcome back to AI Valley.
In today’s edition:
The beginning of OpenAI’s Deep Research
OpenAI has unveiled o3-mini
Figure plans to ship 100K humanoids
UC Berkeley researchers replicate DeepSeek AI for just $30
Plus trending AI tools, posts, and resources
Ready, set, go…
MAIN UPDATE
🤖 The End of Search, The Beginning of OpenAI’s Deep Research
OpenAI has unveiled Deep Research, its next agent that independently conducts multi-step web research on complex topics and delivers detailed reports with citations faster than humans
Here's what you need to know:
The new agent is powered by the upcoming OpenAI o3 model that’s optimized for web browsing and data analysis. It utilizes this model's reasoning capabilities to search, interpret, and analyze massive amounts of text, images, and PDFs on the internet and turn it into a cohesive and well-documented report.
In a demo, the agent generated a comprehensive report on iOS and Android app market trends, showcasing its ability to tackle intricate subjects with accuracy.
The quality of citations are also genuinely advance. Unlike traditional AI-generated sources prone to hallucinations, Deep Research provides legitimate academic references. Clicking a citation often takes users directly to the relevant highlighted text.
To use ChatGPT deep research, simply select “Deep Research” in the ChatGPT composer, enter a query, and optionally attach files or spreadsheets. Reports take 5 to 30 minutes to generate, with a notification upon completion.
This feature closely follows OpenAI’s launch of Operator, its first agent that navigates the web to complete tasks for you. It’s also similar to Google’s Project Mariner, a research prototype that was revealed in December.
The model outperforms all competitors in the Humanity’s Last Exam benchmark, slightly surpassing the o3 preview with a 25.7 score.
Why it matters:
AI is moving closer to general-purpose digital agents: capable of navigating the web, processing multimodal data, and taking meaningful action. While Operator hinted at this future, Deep Research suggests we’re getting closer to AI-powered autonomous research assistants.
TOGETHER WITH NATURA
🧠 Natura introduces AI People—virtual assistants with memory, personality, and initiative
Most AI assistants today feel transactional—resetting with every conversation, losing context, and offering generic responses that never truly evolve.
NatureOS redefines what AI can be.
Created by Natura Umana, it introduces AI People—virtual assistants with memory, personality, and initiative. Unlike traditional chatbots, they remember past conversations, adapt to your preferences, and work together to handle complex tasks. Over time, they become more intuitive, learning from interactions to provide relevant support.
Paired with HumanPods, AI becomes even more seamless. These open-ear AI earbuds let you interact naturally—just tap, speak, and get things done. Designed for all-day comfort, they keep you connected to AI without isolating you from your surroundings. Whether you're working, commuting, or multitasking, AI is always within reach—without the need for screens.
This is more than AI. It’s the future of human-machine collaboration.
Join the waitlist now: join-natura.com/AIV
SIDE UPDATES
OpenAI has introduced o3-mini, a new reasoning model available to both free and paid users. It delivers faster responses and lower costs compared to previous versions, with a 24% speed improvement over o1. The model excels in technical fields like math and coding while being 63% cheaper than o1-mini. Free users now have access to reasoning capabilities for the first time, while paid users get expanded rate limits of up to 150 messages daily.
Figure Founder Brett Adcock made this announcement following the success of a fleet ordered by BMW. The company has recently secured its second commercial customer, described as “one of the largest U.S. companies,” expanding its humanoid robot deployments. This aligns with their aim to address labor shortages in the US by expanding human capabilities through advanced AI. To stay on track for its 100,000-unit target, Figure plans to produce around 1,000 humanoid robots in 2025.
Convergence has recently introduced Proxy, a web-based AI agent designed to tackle complex online tasks with ease, no technical skills required. It navigates websites, forms, and databases much like a human, handling a wide range of tasks on your behalf. It boasts an industry-leading 88% accuracy on the WebVoyager benchmark, outperforming OpenAI’s Operator and Anthropic’s Claude Computer Use. The tool is available globally, offering a free tier for new users and competitive paid plans starting at $20 per month.
Sam Altman, Mark Chen, Kevin Weil, Srinivas Narayanan, Michelle Pokrass, and Hongyu Ren from OpenAI recently held a Reddit Ask Me Anything session to talk about o3-mini and the future of AI. In the thread, the representatives answer questions about advanced voice mode, the upcoming 4o image generator, GPT-5, and more. OpenAI plans to cut API prices in half in August. It doesn't have a timeline for the release of GPT-5 yet - the company is still working on the 4o series.
A team at UC Berkeley, led by PhD candidate Jiayi Pan, claims to have recreated the core technology behind DeepSeek’s R1 Zero model for only $30. Their smaller-scale version, dubbed “TinyZero,” is now available on GitHub for public testing. While it lacks the massive 671-billion-parameter scale of DeepSeek’s original model, Pan says TinyZero successfully replicates key behaviors seen in R1 Zero using reinforcement learning.
TRENDING TOOLS
Voice Coach > AI feedback on interviews, presentations, and conversations.
Gemini AI > It can now help you create charts and uncover data insights about your spreadsheet data.
TalkNotes > A site that turns your ideas into tasks.
Scam AI > Spot scams before they spot you.
Qwen2.5-Max > DeepSeek competitor developed by Alibaba Cloud.
Screenshot-to-code > Take a screenshot of any website and make a clone in minutes.
THINK PIECES / RESOURCES
What fully automated firms will look like.
How to prepare for AGI according to Reid Hoffman.
DeepSeek FAQ by Ben Thompson.
How to train an AI image model on yourself.
How one YouTuber is trying to poison the AI bots stealing her content.
How he is getting $1.69 leads for my friends local coffee shop using ai avatar ugc ads made in 11labs and Heygen.
ImageFX is the Spotify of AI image generators (easy if you struggle with prompts).
AI and two hundred dollar tasks.
AI NPCs try to figure out who among them is the human.
CONTENT CORNER
1/
We have to take the LLMs to school.
When you open any textbook, you'll see three major types of information:
1. Background information / exposition. The meat of the textbook that explains concepts. As you attend over it, your brain is training on that data. This is equivalent… x.com/i/web/status/1…
— Andrej Karpathy (@karpathy)
6:03 PM • Jan 30, 2025
2/
Harvard says AI tutors are better than Harvard professors.
The students who used an AI tutor in a Harvard physics class (instead of the professor’s teaching) learned more than 2X as much.
The best part?
They did it in less time and were more motivated and engaged. x.com/i/web/status/1…
— Austin Scholar (@AustinScholar)
1:11 AM • Jan 22, 2025
3/
4/
THAT’S ALL FOR TODAY
That’s all for today’s issue, folks.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee
👍️ Like what you see? Subscribe here
Thanks for being here.
HOW WAS TODAY'S NEWSLETTER |
REACH 100K+ READERS
Acquire new customers and drive revenue by partnering with us
Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.
If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.