- AI Valley
- Posts
- OpenAI launches o3 and o4-mini 🧠
OpenAI launches o3 and o4-mini 🧠
PLUS: How to vibe code (practical guide)
Together with
Howdy again. It’s Barsee, and welcome back to AI Valley.
Today’s climb through the Valley reveals:
🤝 Inside OpenAI's Controversial Plan to Abandon Its Nonprofit Roots
⚡ Google releases Gemini 2.5 Flash
🧠 OpenAI Launches o3 and o4-mini
🤖 Plus Trending AI Tools, Posts, and Resources
Let’s dive into the Valley of AI…
💌 Top performers get more done with AI-native email
Leading companies are 38% more likely to use an AI-native email app — they don't just settle for Gmail or Outlook.
Superhuman is trusted by the world's fastest-growing companies, including over 60% of the Forbes AI 50.
Fly through your inbox with AI that helps you organize messages, schedule calls, write emails, remember follow-ups, and more.
Emails that write themselves. When it's time to follow up, Superhuman will remind you and draft the email — all you have to do is send.
A more focused inbox. Superhuman can automatically archive cold pitches, marketing emails, and more via Auto Labels. Create your own labels with AI prompts. Eliminate inbox clutter for good.
AI that sounds like you. Superhuman AI adapts its writing style to the person you're emailing and to your unique voice and tone.
*This is sponsored
THROUGH THE VALLEY
OpenAI, now valued at $300 billion, operates as a for-profit company under a nonprofit board. After Sam Altman’s firing in 2023, investors pushed to drop its hybrid structure. The firm has long struggled to balance fundraising and its mission, but recent moves aim to separate the two. The board will finalize OpenAI’s future and redefine its nonprofit role before 2025. This article explores the challenges behind this major transition.
OpenAI has quietly rolled out "Memory with Search," an enhancement to ChatGPT that allows the AI to incorporate details from your past conversations when conducting web searches. This feature builds upon ChatGPT's recently expanded memory capabilities, enabling more personalized search results based on your previously shared preferences and information.
Google’s new Gemini 2.5 Flash is a hybrid AI model rivaling o4 mini and beating Claude 3.5 Sonnet in reasoning and STEM tasks at lower costs. It introduces a "thinking budget" feature, letting developers adjust processing up to 24k tokens for optimal speed and quality. Currently in preview, it’s available via Google AI Studio, Vertex AI, and the Gemini app, offering a cost-efficient alternative for advanced AI applications.
At TED 2025, OpenAI CEO Sam Altman discussed the company’s rapid growth to 800 million weekly users and the infrastructure strain from high demand. He addressed OpenAI’s shift from a nonprofit to a $300 billion giant and growing concerns over AI’s societal risks, including power concentration and autonomous agents. Altman acknowledged critiques while highlighting the challenges of scaling responsibly amid intense scrutiny.

Beijing hosted a half-marathon featuring 21 Chinese humanoid robots, showcasing the country’s robotics ambitions. The fastest bot finished in 2:40:42—far behind the human winner’s 1:02:36. China aims to lead in humanoids by 2027, offering subsidies, tax breaks, and talent incentives to boost the sector. The event demonstrated progress but also highlighted the gap between machine and human physical capabilities.
ChatGPT’s latest models, o3 and o4-mini, can now pinpoint locations from photos, without metadata, raising privacy concerns. Users are testing its GeoGuessr-like abilities, identifying landmarks and cities from visual clues alone. While impressive, the AI sometimes fails, revealing its limits. The feature underscores how AI can extract private details from seemingly harmless images, blurring the line between fun and surveillance in an increasingly data-driven world.
Tamay Besiroglu, co-founder of Epoch, has launched Mechanize, an AI startup targeting full automation of human jobs. Initially focusing on white-collar tasks like data entry and research, the firm trains AI agents in virtual environments to handle complex, long-term work. The ambitious goal? A future where AI performs all jobs, adapting to challenges without human intervention, sparking debates over employment and economic disruption.
PEAK OF THE DAY
🧠 OpenAI launches o3 and o4-mini
o3 really blew my mind with this one.
I gave it an image of a menu of my favorite Chinese place in SF with no title or EXIF data, and it was able to search the web, match menu items, and locate it.
🤯
— Deedy (@deedydas)
8:42 PM • Apr 16, 2025
OpenAI has introduced o3 and o4-mini, its most advanced reasoning models yet, capable of integrating images directly into their "chain of thought" and utilizing all ChatGPT tools with full agentic access.
What's new?
Image as part of reasoning: Unlike older models that could only "see" images, o3 and o4-mini can reason with them. They can rotate, zoom, transform, and use visuals dynamically to solve problems.
Understands low-quality images: You can upload whiteboards, notes, or rough sketches, and the models will accurately analyze even messy visuals.
Independent tool use: The models can autonomously use all ChatGPT tools, such as browsing, Python, image generation, and analysis, to tackle multi-step problems.
Rigorous safety testing: OpenAI says these models went through their most intense safety program yet, based on the recently updated "Preparedness Framework.”
How good are they?
o3 sets new SOTA performance across coding, math, science, and multimodal benchmarks. It also makes 20% fewer major mistakes in real-world workflows like programming, consulting, and creative ideation.
o4-mini offers fast, cost-efficient reasoning, significantly outperforming previous mini models and even beating o3 in benchmarks like AIME 2025 math.
What made this leap possible?
OpenAI used 10x more training compute compared to o1, pushing their model capabilities much further.
Is there any issue?
Both models tend to hallucinate more than previous ones. o3 hallucinated 33% of the time on the PersonQA benchmark, and o4-mini did even worse at 48%. Third-party tests have also confirmed these issues, with o3 sometimes fabricating actions it supposedly took.
How to access them?
They are already live for Plus, Pro, and Team users in ChatGPT, replacing earlier versions. Enterprise and Edu users will get access soon.
Why does it matter?
These models aren't just better at answering questions; they're becoming independent problem-solvers. By combining reasoning, tools, and image manipulation, o3 and o4-mini mark a step closer to AI that collaborates more like a teammate than a chatbot.
TRENDING TOOLS
Happenstance > Quickly discover key contacts across social networks using plain English.
Gemini AI Video Generator > Describe what you have in mind and watch your ideas come to life in motion.
Backed AI > Your Daily AI Companion for Back Pain & Posture Correction.
Claude Research > Claude takes research to new places.
Omakase Voice > Turn your website into a voice-powered sales agent.
Infinite Reality > Create interactive 3D websites and virtual experiences for your brand without coding.
Codex CLI > OpenAI’s open-source AI coding assistant that helps you write and edit code directly from your terminal.
THINK PIECES / BRAIN BOOST
How to vibe code (practical guide).
AI Engineering in 76 minutes (Complete Course/Speedrun!)
GPT 4.1 Prompting Guide by OpenAI Cookbook.
Always set the temperature to zero if you want the model to stick to the prompt.
The social distribution of wealth.
How we build effective agents: Barry Zhang, Anthropic.
Vibe Coding is not an excuse for low-quality work.
A realistic AI timeline.
AI is like cars.
A practical guide on building AI agents from OpenAI.
ChatGPT spends 'tens of millions of dollars' on 'please' and 'thank you'.
VALLEY GEMS
1/
we’ll look back at this era like the gold rush.
except this time:
– picks + shovels = prompts + AI agents
– gold = attention, data, distribution
– miners = builders automating boring work
– gold pans = n8n, replit, bolt, lovable
– land grabs = ai-first domains + keywords
–— GREG ISENBERG (@gregisenberg)
4:37 PM • Apr 17, 2025
2/
Y Combinator CEO Garry Tan has said that for about a quarter of the current YC startups, 95% of the code was written by AI.
— unusual_whales (@unusual_whales)
3:37 PM • Apr 16, 2025
3/
Bill Gates is warning that all doctors and teachers will be replaced by AI in less than 10 years.
— Financelot (@FinanceLancelot)
2:16 PM • Apr 20, 2025
4/
Major models released in last 4 months:
DeepSeek R1
o3-mini
Qwen 2.5-Max
Grok 3
Grok 3 mini
Claude 3.7
QwQ-Max
Gemini 2.0 Flash
Gemini 2.0 Pro
GPT-4.5
Gemini 2.5 Pro
Llama 4 Scout
Llama 4 Maverick
GPT 4.1
GPT 4.1-mini
GPT 4.1-nano
o3
o4-mini
Gemini 2.5 Flash— Theo - t3.gg (@theo)
8:29 PM • Apr 19, 2025
5/
Vibe coding is where mobile devices were in 2003 pre-iPhone.
GPRS was crap and unreliable
The screens were tiny
Almost no websites really worked in mobile browsers
But boy did that change a few years later
— Garry Tan (@garrytan)
5:23 PM • Apr 20, 2025
THAT’S ALL FOR TODAY
Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee
👍️ New reader? 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”.