- AI Valley
- Posts
- DeepSeek Unveils Prover V2
DeepSeek Unveils Prover V2
PLUS: Meta AI app debuts with discover feed to take on ChatGPT
Together with
Howdy again. It’s Barsee, and welcome back to AI Valley.
Today’s climb through the Valley reveals:
➗ DeepSeek unveils Prover V2
🦙 Meta AI app debuts with discover feed to take on ChatGPT
🧠 $2B deal gives Mira Murati unshakable control over AI startup
🤖 Plus trending AI tools, posts, and resources
Let’s dive into the Valley of AI…
The new Advanced Style Creation and Control tools give creative teams the power to build, remix, and maintain standout visual styles — all with AI.
Explore endless styles (vector art, icons, logos, 3d images, illustrations, etc)
Tweak and remix until it’s perfectly you
Share styles across teams to keep things consistent
It’s brand identity, supercharged.
Try it free — and unlock full access for just $1 with code VALLEY11
*This is sponsored
THROUGH THE VALLEY
🦙 Meta AI App Debuts With Discover Feed to Take on ChatGPT
Meta just launched its standalone AI app, powered by Llama 4, to compete with ChatGPT. You can chat, generate images, and get web results, but the real twist is its social "Discover" feed, where users share AI interactions like social posts. Friends can like, comment, or remix them. The app also pushes voice chats, with an experimental "full-duplex" mode for natural conversations. While already used by 1B+ people in Meta’s apps, this dedicated version merges with its Ray-Ban smart glasses companion app, hinting at future AI-hardware fusion.
🚘️ Waymo, Toyota strike partnership to bring self-driving tech to personal vehicles
Waymo (Google’s self-driving division) and Toyota are exploring a partnership to bring robotaxi tech to personal cars. This could mean Toyota vehicles joining Waymo’s ride-hailing fleet. Waymo already operates 250,000 paid rides weekly in cities like San Francisco, LA, Phoenix, and Austin, but this deal could bring autonomy to privately owned cars.
🤖 Xiaomi Enters China's AI Race with MiMo Reasoning Models
Smartphone giant Xiaomi has launched MiMo, a 7B-parameter open-source AI model for math, coding, and reasoning. It comes in four versions, some fine-tuned for better performance. Xiaomi claims it beats OpenAI’s o1-mini and Alibaba’s Qwen-32B, though independent tests are pending. You can try it now on Hugging Face and GitHub.
🧠 $2B Deal Gives Mira Murati Unshakable Control Over AI Startup
Mira Murati, former OpenAI CTO, is raising 2 billion for her new AI startup, Thinking Machines Lab —valuing it at 10B. The deal ensures she always has majority control, no matter who joins the board later. Despite having no product or revenue yet, big investors like Andreessen Horowitz (a16z) backed her, believing in her long-term vision for Artificial General Intelligence (AGI).
🚨 Musk Teases Grok 3.5 Launch for Early Beta
Elon Musk just revealed that Grok 3.5, xAI’s next-gen model, will enter early beta next week, but only for SuperGrok subscribers. Unlike most AI models, Grok 3.5 doesn’t rely on internet data—instead, it reasons from first principles, making it better at solving complex technical questions.
💳 Mastercard Gives AI Agents the Ability to Shop on Your Behalf
Mastercard just introduced Agent Pay, a system that allows AI assistants (like Copilot) to search, compare, and buy products on your behalf. It uses secure Agentic Tokens to keep payments safe while giving users full control. Partners like Microsoft and IBM are already working to expand this AI-powered shopping tech.
PEAK OF THE DAY
➗ DeepSeek unveils Prover V2: An open-source math powerhouse
DeepSeek has quietly launched Prover V2, an open-source model built to solve math problems using Lean 4 assistant, which ensures every step of a proof is rigorously verified.
What's impressive about it?
Massive scale: Based on DeepSeek-V3 with 671B parameters using a mixture-of-experts (MoE) architecture, which activates only parts of the model at a time to reduce compute costs.
Theorem solving: Uses long context windows (32K+ tokens) to generate detailed, step-by-step formal proofs for a wide range of math problems — from basic algebra to advanced calculus theorems.
Research grade: Assists mathematicians in testing new theorems automatically and helps students understand formal logic by generating both Lean 4 code and readable explanations.
New benchmark: Introduces ProverBench, a new 325-question benchmark set featuring problems from recent AIME exams and curated academic sources to evaluate mathematical reasoning.
How good is it?
It achieves state-of-the-art results in neural theorem proving, with an 88.9% pass rate on the MiniF2F-test (high school level), beating GPT-4 and Claude. It also solves 49 out of 658 problems from PutnamBench (undergraduate level), outpacing most open-source models.
Can I run it on my laptop?
Yes, its 7B version can run on laptops with 24GB VRAM thanks to FP8 quantization, which compresses the model without losing much accuracy.
Where can I try it?
Both the 7B and 671B versions are available on Hugging Face with MIT licenses, allowing free commercial use, modification, and integration into other tools.
Why does it matter?
This is a huge leap in AI-driven mathematics. DeepSeek Prover V2 isn't just solving problems, it's rigorously proving them, which could redefine how we explore, teach, and verify math.
TRENDING TOOLS
OpenNote > Build a personalized learning space with videos, whiteboards, graphs, and practice problems that adapt to you.
Hana by Spline AI > A real-time 2D design canvas to create and animate interactive, code-free interfaces and visuals with ease.
Promptimize > One-click prompt enhancement for expert-level results across all AI platforms.
Hera > Turns your text into stunning motion graphics instantly.
Adobe Firefly 4 > Generate images, video, audio, and vectors in one platform with fast & ultra-realistic models.
THINK PIECES / BRAIN BOOST
Avoiding skill atrophy in the age of AI.
Ask HN: AI replacing engineers – firsthand stories?
The Prompt is Becoming the Product by Nate B Jones.
I Tried Vibe Coding: How I Created GTA with Artificial Intelligence by Tomáš Bencko.
"Vibe Coding" isn't new.
Neuromorphic computing: the future of AI.
VALLEY GEMS
Meta CEO, Mark Zuckerberg:
"within 12-18 months, most of the code is written by AI"
It won’t just be autocomplete.
AI agents will set goals, run tests, find problems, and write better code than top engineers.
— Haider. (@slow_developer)
5:07 PM • Apr 29, 2025
What Sam Altman wishes someone had told him:
— Ben Lang (@benln)
5:10 PM • Apr 30, 2025
New 42 min deep dive: Using Claude, Gemini & Grok to build a 3D city simulation, dynamic video annotations, smart shot lists, and even a AR HUD overlay. Chapters below:
00:00 LLM Superpowers You're Not Using
00:26 Claude 3.7 Overview
02:38 3D City Vibe Coding for Creatives— Bilawal Sidhu (@bilawalsidhu)
2:40 PM • Mar 7, 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”.