Good Morning! Chinese AI lab DeepSeek just dropped preview versions of its highly anticipated V4 models, claiming the title of the largest open-weight model in the world while brutally undercutting Western competitors on price. Plus, I’ll show you how to isolate vocals in a song with Google Flow Music AI.
DeepSeek V4 "Closes the Gap" With Frontier Models
ChatGPT Solves 60-Year-Old Math Problem
OpenAI Drops 5-Principle AGI Framework
How to Isolate Vocals in a Song with Google Flow Music AI
4 new AI tools worth trying
AI MODELS
DeepSeek has launched V4 Flash and V4 Pro. Both are mixture-of-experts (MoE) models boasting massive 1-million-token context windows, designed to aggressively compete with the latest frontier releases from OpenAI, Google, and Anthropic.
The V4 Pro is a behemoth with 1.6 trillion parameters (49 billion active), officially making it the largest open-weight model available today. The smaller Flash model has 284 billion parameters (13 billion active).
DeepSeek claims V4 Pro outperforms its open-source peers in reasoning, beats GPT-5.2 and Gemini 3.0 Pro on certain tasks, and practically matches GPT-5.4 in coding benchmarks.
Pricing is insanely aggressive: V4 Flash costs just $0.14 per million input tokens, massively undercutting GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5.
The models are currently text-only and lag about 3 to 6 months behind state-of-the-art frontier models like GPT-5.4 and Gemini 3.1 Pro in general knowledge tests.
The release arrives amid massive geopolitical drama, launching just one day after the U.S. accused China of stealing American AI IP on an industrial scale using thousands of proxy accounts.
DeepSeek is proving that China's AI ecosystem can keep pace with Silicon Valley despite strict hardware export bans. By offering a 1.6-trillion-parameter open-weight model at a fraction of the cost of Western frontier models, DeepSeek is applying immense pricing pressure on OpenAI and Anthropic, forcing a race to the bottom for inference costs while reigniting massive tensions over AI IP theft and model distillation.
AI MODELS
Liam Price, a 23-year-old without formal mathematical training, used a single prompt in ChatGPT-5.4 Pro to solve one of the infamous "Erdős problems" regarding "primitive sets," completely bypassing a mental block that had trapped human experts for 60 years.
Erdős problems are a series of legendary math conjectures left behind by prolific mathematician Paul Erdős. This specific problem involved calculating the score limits of primitive sets (sets of whole numbers where no number can be evenly divided by another).
Unlike previous AI math solutions that simply regurgitated known methods, ChatGPT-5.4 Pro took an entirely novel route, applying a formula from an adjacent subfield that human experts simply hadn't thought to use.
The raw output of the AI was "quite poor" and required top mathematicians like Stanford's Jared Lichtman and UCLA's Terence Tao to sift through the slop, but they confirmed the core insight was a genuine, novel cognitive leap.
Price and his collaborator previously sparked this trend by feeding random open problems to LLMs just to see what happened, a process researchers are now affectionately calling "vibe mathing."
We are seeing the transition from "vibe coding" software to "vibe mathing" advanced academia. When a 23-year-old with a $20/month AI subscription can solve a 60-year-old problem that stumped Stanford and Cambridge PhDs, the barrier to entry for world-class discovery has fundamentally shifted. The AI didn't just calculate faster; it successfully reasoned its way around a human mental block, proving that LLMs can generate truly original scientific insights.
AGI
OpenAI has published a new five-principle framework outlining its approach to artificial general intelligence, serving as its most prominent statement of intent since its foundational 2018 Charter.
The framework is built on five core pillars: democratization, empowerment, universal prosperity, resilience, and adaptability.
CEO Sam Altman explicitly pledged to resist concentrating AI power in the hands of the few, committing to keeping AGI decisions subject to "democratic processes" rather than leaving them strictly up to the AI labs.
Altman made an unusual concession regarding OpenAI's massive scale today, acknowledging that the company may have to trade off some user "empowerment" in exchange for more "resilience" (safety and security).
The post reflects on OpenAI's infamous 2019 decision to delay the release of GPT-2, calling the panic "misplaced" but crediting the episode for establishing the "iterative deployment" strategy they still use today.
This policy drop comes while OpenAI is heavily on the defensive, following major public backlash over their recent Pentagon deal and tightening regulatory oversight in California and Europe.
OpenAI is playing defense. After months of intense scrutiny over military contracts and a wave of new state and federal AI laws, the company is trying to control the narrative and set its own public posture before regulators forcefully write the rules for them. Furthermore, Altman's rare concession that OpenAI will trade "empowerment for resilience" is a clear signal: as these frontier models get closer to actual AGI, expect them to become much more strictly guarded and controlled.
HOW TO AI
🗂️ How to Use Advanced Audio Editing in Google Flow Music
In this tutorial, you will learn how to use the latest AI-driven features in Google Flow Music to isolate vocal tracks from instruments and apply professional-grade audio effects using simple natural language commands.
🧰 Who is This For
Music producers and beatmakers refining tracks
Podcasters cleaning up voice recordings
YouTubers editing videos with better audio
Content creators improving sound quality
Sound designers crafting effects and layers
STEP 1: Isolate Vocals and Instruments (Creating Stems)
Load your generated song or upload an audio file into Google Flow Music. To separate the tracks, simply type a prompt like: "Create stems from this music" or "Isolate the vocal from the instruments."
The AI will process the file and present you with four primary labels: Drums, Bass, Vocals, and Other (which includes guitars or pianos). Use the "M" (Mute) button to silence a specific track or the "S" (Solo) button to listen to one track by itself.
STEP 2: Download Your Separated Tracks
Once you are happy with the isolation, click the Download button. Google Flow Music will generate a ZIP file containing each individual audio track as a separate file. This is perfect for remixing in external software or using just the instrumental version for a project.
Note that if a song doesn't contain a specific element (like a bassline), that file may appear empty or be excluded from the ZIP.
STEP 3: Apply Fast Audio Effects
You can modify the "vibe" of your music without re-generating the entire song using the Audio Effects feature. This works like an intelligent equalizer and is significantly faster than standard AI generation. Use a descriptive prompt such as:
"Apply audio effects to make the music sound like it is being played in a hotel lounge" or "Make the vocals sound nostalgic and echoey."
The AI will instantly adjust the reverb, delay, and gain settings to match your description.
STEP 4: Fine-Tune with the FX Dashboard
After applying an effect, click the FX button on the specific track (Vocal, Drums, etc.) to see the exact settings the AI applied. Even if you aren't familiar with technical numbers for things like "Chorus" or "Reverb," you can see the workflow the AI used.
You can apply these effects to the whole song or micro-target them to just the vocals while keeping the instruments clear, giving you total creative control over the final mix.
ASML says it plans to make at least 60 of its standard EUV machines this year, 36% more than it sold in 2025, as it races to meet demand for making AI chips.
Anthropic published a post-mortem tracing Claude Code quality complaints to three separate bugs, resetting usage limits for subscribers due to the issues.
Microsoft says Copilot's agentic features in Word, Excel, and PowerPoint are now generally available and enabled by default for 365 Copilot and 365 Premium.
Tokyo-based humanoid robotics startup Genki Robotics, co-founded by Andy Rubin, raised a Series A at a ~$1B valuation; it raised a ~$50M seed in 2025.
🌎 DeepSeek V4:The lrgest open-weight model in the world
🎆 ChatGPT Images 2.0: OpenAI’s upgraded image generator
🌎 GPT5.5: OpenAI’s most powerful AI model built using recursive self-improvement.
💻 Holo 3: Open AI agent that can use computers like a human
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas