DeepSeek Got a Massive Upgrade (v3.1)!

Dear Readers,

Anyone who believes that open source has no chance in the race for the best AI models should take a look at DeepSeek v3.1: two modes, 128K context length, and a fraction of the cost—suddenly, what was previously only possible for billion-dollar corporations is now within reach. It is precisely this shift in the balance of power that makes the AI market so exciting right now.

In this issue, we take you through the biggest leaps of the week: from DeepSeek and its challenge to GPT-5 to new antibiotics designed by AI that could fight super bugs. Also: a protein that reverses brain aging and multimodal models that detect Alzheimer's markers without expensive PET scans. Each of these developments shows how closely technology and life extension are now intertwined – and how close the future really is.

In Today’s Issue:

All the best,

DeepSeek-v3.1 takes the race for efficient AI models to a new level. The model was developed to support two modes: a “thinking mode,” which works with extensive internal deliberation, and a “non-thinking mode” for fast, resource-efficient responses. With a context length of 128,000 tokens, it can handle significantly more complex tasks without losing track. Particularly interesting: thanks to the FP8 data format and optimized architecture, it is much more cost-effective to operate and requires significantly less computing power than comparable models.

Compared to the previous version 3.0, the current version is a huge leap forward and demonstrates outstanding performance in numerous benchmarks. In short, version 3.1 is a massive upgrade over its predecessor.

For the AI community, this means powerful agent capabilities and an API that can be easily integrated into existing applications without breaking the budget. DeepSeek thus demonstrates that open-source models can now compete with closed solutions such as GPT-4 or Claude in terms of quality and efficiency – at a fraction of the cost. The bigger question now is what new applications will become possible when AI is no longer held back by high infrastructure costs.

Why it matters: DeepSeek-v3.1 makes advanced AI affordable and lowers the barriers to entry. This brings broader use in research, business, and everyday applications closer.

Sources:

🔗 https://huggingface.co/deepseek-ai/DeepSeek-V3.1

🔗 https://www.reuters.com/world/china/chinese-startup-deepseek-releases-upgraded-ai-model-2025-08-21/

🔗 https://venturebeat.com/ai/deepseek-v3-1-just-dropped-and-it-might-be-the-most-powerful-open-ai-yet/

🔗 https://en.wikipedia.org/wiki/DeepSeek

How do you follow record-setting success? Get stronger. Take Pacaso. Their real estate co-ownership tech set records in Paris and London in 2024. No surprise. Coldwell Banker says 40% of wealthy Americans plan to buy abroad within a year. So adding 10+ new international destinations, including three in Italy, is big. They even reserved the Nasdaq ticker PCSO.

Invest for $2.90/Share

_{Paid advertisement for Pacaso’s Regulation A offering. Read the offering circular at invest.pacaso.com. Reserving a ticker symbol is not a guarantee that the company will go public. Listing on the NASDAQ is subject to approvals.}

According to a beta tester, Google's new "Co-Scientist" AI is "unbelievably good" at generating novel ideas, hypotheses, and scientific proposals, and will fundamentally change how science is done.

MIT researchers have designed two novel antibiotics—NG1 (against resistant gonorrhea) and DN1 (also against MRSA). The models searched through ~36 million chemical possibilities; the lead molecules were effective in cell tests and mouse models (Cell, Aug. 14, 2025). New: De novo design instead of re-screening known substances. Relevance: Faster supply of drugs against superbugs. Outlook: Optimization and several years of clinical trials are pending.

UCSF researchers identify the iron-associated protein FTL1 as a driver of cognitive aging in the hippocampus of mice in Nature Aging. FTL1 is elevated in old age; overexpression makes young animals “old,” targeted reduction restores synapse function and memory; metabolic support (including NADH) mitigates effects. New: a single, causal target. Relevance: potential approach for more precise drug strategies against cognitive decline.

A Nature Portfolio team combines Robust PCA with transformer networks and reduces the error in age predictions from skin, mouth, and gut microbiomes by up to 28% (≈18,000 samples from 60+ studies). New: strong, cross-dataset generalization. Relevance: Microbiome clocks are becoming more robust and non-invasive. Future: More accurate “pace of aging” trackers and personalized nutrition/probiotics instead of simply counting years.

Working at the intersection of AI and human longevity? We’re inviting researchers and innovators to submit their work to be featured in Superintelligence, the leading AI newsletter with 200k+ readers.

Whether you’re using AI to extend lifespan, improve healthspan, or accelerate biomedical breakthroughs, email your arXiv.org link or a summary to [email protected] with the subject line “Longevity Submission”. If selected, we’ll contact you for a potential feature.

xAI fixed a bug that has been on X/Twitter for 18 years.