Deepseek's New Model: R1052A
Deepseek has released a new model, R1052A, which is claimed to reason deeper than anything they've ever shipped. It performs well on their own chart, almost matching OpenAI 3, and even surpasses Google's Gemini 2.5 Pro in some tests. The model has been trained on logic, puzzles, math, and coding, and Deepseek has been honest about the issue of hallucinations. The interface remains the same, but the web search has improved. The model can be downloaded and run on a decent laptop, providing a private LM option. However, it's too early to crown it as perfect, and real tests will come when it's applied to everyday tasks.
Web Search and Proxies
The web search in Deepseek didn't work before, possibly due to the way it was scraping the web for data. Sites usually block requests from the same IP address. The simple fix is to use clean, high-quality proxies. Node Maven is a sponsor of the video and offers over 90% clean and safe IPs, compared to around 25% guaranteed by the biggest brands. Their proxies are customizable, and you pay only for data, not for proxies. Unused bandwidth never expires.
Deepseek Prompts and Rules
When using Deepseek, treat the first message like a set of house rules. Spell out exactly what you want, the language you expect the answer in (English or Mandarin), and any other requirements. If the project is complex, provide a road map. Details are important, and you can ask the model to grade itself and point out weak spots. You can also run the whole model on your own laptop without relying on the cloud or cafe Wi-Fi.
Deepseek R1 Model
The Deep Sea Crew has squeezed their new R1 model down to a carry-on edition that can run on a single GPU with 8 billion parameters. It still outperforms Google's Gemini 2.5 Flash in some benchmarks. The model borrows the skeleton from Alibaba's 8 bill Quan 3. It can be fired up on a big boy card with 40 to 80 gigs of VRAM. The new version is a playground for researchers and a launchpad for small industrial gigs, offering solid muscle without the need for a server farm.
Censorship and Bias
All the new models have a "made in China" sticker, which brings along the issue of censorship. Deep Seek claims to be independent, but the model still has limitations. It won't touch contentious free speech topics like interment camps or criticize Beijing. The model lives in Hugging Face, and as more people use it, bias can drift in. The safest move is to verify everything and treat every paragraph like a bar story.
OpenAI and Wix Collaboration
OpenAI has collaborated with Wix to build websites. Instead of dragging boxes around, you can use a chat to create a finished site with layout, images, copy, and even booking or store apps. GBT can write blog posts, edit or generate header pics, spit out business reports, and translate the site into other languages. This could make it easier for people to launch their own personal sites.
Meta AI's 1 Billion Monthly Active Users
Meta AI claims to have 1 billion monthly active users across all their apps, but this includes people who may have never used the chatbot. Zuk is counting installations, not chats, which is more of a marketing spin than a real milestone. Facebook has been struggling, and WhatsApp still lacks some basic features.
Perplexity's Upgrade
Perplexity has released a fat upgrade for pro users called Labs. It allows you to spin up AI projects, which are mini apps that can perform multi-step jobs. You can also use it as a Swiss Army analyst for tasks like drafting marketing campaigns and untangling P&L statements. Perplexity has also added deep research, which provides long citation-packed briefs.
Hugging Face's Humanoid Robots
Hugging Face has unveiled two humanoid robots, a tall gangly bot with 66 joints and a smaller sidekick. Both robots are open source, and the price is relatively affordable. Hugging Face wants robotics to be accessible to everyone, not just a few tech giants.
Telegram and XAI Collaboration
Telegram has partnered with Elon's XAI to bring Grock inside the app. XAI is investing around 300 million dollars in cash and equity, and Telegram will get half of every Grock subscription that runs through their app. Grock will be able to polish messages, summarize chats, digest links and docs, and create custom stickers. However, privacy is a concern.
Opera's Neon Browser
Opera is launching a new browser called Neon with baked-in AI workflows. It can shop for you, autofill forms, and even write code. The browser ships with a personal agent under the hood. However, it's currently behind a wait list, and there may be a subscription fee.
Conclusion
The AI scene is full of exciting developments, with new models and collaborations emerging every week. While there are some concerns about censorship and bias, the potential benefits of AI are significant. It's important to stay informed and use AI tools responsibly. If you're interested in learning more about AI, check out the Geek Academy link in the description.