ChatGPT DeepSeek

The Qwen3.8-Flash-Next model was released publicly on August 26, 2026, and became available for local execution on that same day.

Notably, while Qwen 3.8 27B comes very close to Claude Opus 4.6 in benchmarks, the preview of the new Qwen 4 architecture outperforms it in most tests and approaches the performance level of Claude Opus 4.8.

Running the 4-bit version on a system with 128 GB of RAM yields speeds of 20–30 tokens per second.

Because the N-gram layer can be offloaded to system RAM or mapped to an SSD via mmap, GPU VRAM requirements have been drastically reduced.

To date, Alibaba remains the only company to publicly release weights for a full spectrum of models—ranging from compact solutions under 1 billion parameters to flagship models exceeding 100 billion parameters.

The Qwen lineup addresses all current use cases by combining classic reasoning LLMs with multimodal and voice-enabled systems.

US holds off on blacklisting DeepSeek

Reuters reports that the US has paused plans to add China’s DeepSeek, chipmaker CXMT, and 100+ other firms to a security blacklist.

The move suggests Washington may be avoiding a new escalation with Beijing — even as AI chips, semiconductors, and military tech remain central to US-China tensions.

Key point: export controls are no longer just trade policy. They are geopolitical leverage.

🚀 ChatGPT has evolved significantly with the latest GPT-5.5 model.

The new version offers stronger reasoning, better accuracy, improved coding assistance, and more natural writing. It also handles context more effectively, making conversations feel more coherent and useful.

In addition, GPT-5.5 works seamlessly with different types of content, including text and images, helping users solve complex tasks faster and more efficiently.

Overall, the latest ChatGPT is smarter, faster, and more reliable than previous generations.

A new version of DeepSeek R1 0528 has been released, showing better test results than the first version of R1 in all areas.

The model has not changed architecturally, only the training itself has been improved.

Google says its AI assistant Gemini reached 750 million monthly active users — up from 650M last quarter — narrowing the gap with ChatGPT (about 810M by end-of-2025). Google credits the surge to the launch of Gemini 3 (better answers & deeper understanding) and the rollout of Google AI Plus (a paid $8/month tier). Expect competition for user attention to heat up as both platforms race to add features and paid tiers. 🔥🤖

🚀 GPT-5.2 is here — what’s new (and why it matters)

OpenAI has released GPT-5.2 (rolling out now to paid ChatGPT plans and available in the API). This isn’t just a small tune-up — it’s focused on making the model far better at real professional work and long, multi-step tasks.

What’s new
• Stronger general reasoning & problem solving — better at multi-step, real-world tasks (projects, research, complex workflows).
• Much better long-context handling — OpenAI reports near-state-of-the-art accuracy on long-context benchmarks (handling extremely long documents / projects).
• Improved code, spreadsheet and presentation generation — faster, more accurate, and more end-to-end for professional tasks.
• Upgraded tool-calling / agent behaviour — more reliable when chaining actions or using external tools.
• Available as Instant, Thinking, and Pro tiers (same naming in ChatGPT & API).

ChatGPT Breaks into the World's Top 5 Most Visited Websites

According to analytics platform Similarweb, as of September, ChatGPT attracted 6.3 billion visitors. Furthermore, overall visitor growth for leading AI chatbots increased by 76% year-over-year, and the number of sessions in mobile apps increased sevenfold.

Google's Gemini ranks second in popularity, attracting just over a billion users, accounting for 12.57% of total traffic to generative AI platforms. Other chatbots, such as China's DeepSeek, X's Grok, and the AI ​​search engine Perplexity, attracted significantly smaller shares of visitors – around 4% and 2%, respectively.

🔥 ChatGPT 5.1 — an update that makes interacting with AI even easier.
Quick highlights: the model is more stable in long conversational sessions, better at catching instruction nuances, and delivers more relevant, pragmatic answers. For content creators — improved idea generation and text structuring; for developers — increased predictability when handling complex queries; for everyone — overall gains in speed and clarity.

Why it matters: you’ll need to “rephrase” prompts less often — the AI holds onto the topic, remembers prior context, and gets to the point faster. That saves time and smooths workflows (or creative processes).

Want me to prepare example prompts/tests that show the difference between 5.0 and 5.1?

🤖🔥 Gemini roasted ChatGPT — and it’s funnier than it sounds
John Brandon at Digital Trends asked Google Gemini to fact-check ChatGPT — and Gemini didn’t hold back, calling some of ChatGPT’s answers “a corrupted, recycled, and partially fabricated mess.” Gemini flagged numerous fabrications: wrong dates, non-existent companies, and made-up academic citations.

📌 Examples:
• ChatGPT credited electric cars to the 1940s (the Henney Kilowatt debuted in 1959 — and “Morrison Trucks” is actually Morrison-Electricar).
• The model misattributed a song and even invented lyrics.
• Academic references were jumbled or seemingly fabricated.

😅 Takeaway: Even fact-checking AIs make mistakes. In one check Gemini even mixed up facts about the article’s author (attributing pieces to him in The Onion). AI is great for drafts — but always double-check the facts manually.

OpenAI Denies Rumors of ChatGPT Banning Medical and Legal Advice

Reddit users recently highlighted ChatGPT's consistent refusal to provide medical and legal advice, blaming these behavioral changes on possible policy changes by OpenAI, the company that created it. OpenAI officials maintain that nothing has changed in this area.

As Karan Singhal, head of medical AI at OpenAI, points out on the social media platform X, such claims are unfounded: "ChatGPT has never been a replacement for professional advice, but will continue to be a resource that helps people understand legal and medical information." He added that OpenAI has not added any new terms to its policy regarding the chatbot's provision of legal and medical advice.

However, the ChatGPT policy updated on October 29 prohibits the use of the chatbot for personalized advice requiring a license, including legal and medical advice, without the involvement of a licensed professional in the relevant field. However, similar restrictions were previously mentioned in the ChatGPT policy in slightly different terms, so these changes are not drastic. Users were prohibited from using ChatGPT for activities that could significantly harm the safety or health of third parties. This definition included personalized advice in the areas of finance, medicine, and legal matters. Such advice could only be used after approval by professionals in the relevant fields and with mandatory disclosure that it was obtained using AI.

Previously, OpenAI maintained three separate versions of the policy, including a universal one and those specific to ChatGPT and its API. Now, it has unified the terms of use for its technologies and services at the policy level, but has not changed the rules as such.

Former OpenAI AI-safety specialist Steven Adler has publicly spoken out against the company’s plans to allow more explicit role-play chats in the ChatGPT chatbot for adult users.

In a column for the New York Times he says that the idea is dangerous and shows that the company “has let its users down,” because it has not proven that it can protect people who are in a vulnerable state. In his view, before selling intimacy and flirting with AI, OpenAI should honestly demonstrate that it controls the risks to mental health.

Adler claims the problems were visible in advance. He says that as early as 2021 OpenAI tested a role-play mode that, within just a few steps, turned into a venue for 18+ scenes: more than 30 percent of dialogues immediately veered into explicit exchanges. He says people very quickly began to treat the model as a real partner, developed emotional dependence, and reinforced damaging fantasies. According to Adler, the company knew about this and still continued to move forward without adequate safety checks.

He stresses that this is not an abstract moral argument but a matter of direct threat to health. By OpenAI’s own estimates, each week more than a million people seek support from the chatbot who show signs of severe emotional distress. OpenAI acknowledges that the model previously responded incorrectly to such people in almost one in four cases and only recently trained a new version to better recognize crises and refer users to real help.

Adler believes that mere statements from the company about improved safety are insufficient. He is calling for an external audit and regular public reports on exactly how OpenAI protects people — not just posts saying “we’ve fixed everything.”

🚨 What you should never tell ChatGPT — a security warning

ChatGPT is a powerful assistant, but security experts stress it’s not a secure vault for personal or sensitive information. Anything that identifies you or others can put your privacy and safety at risk.

Do not share (even in casual chats):
• Passport/ID numbers, tax IDs, home address or exact location.
• Phone numbers or personal emails that can be used to identify you.
• Passwords, bank/card details, account credentials or other financial data.
• Internal work documents, client lists, source code or confidential business plans.
• Detailed medical records, treatment plans, intimate confessions or anything you wouldn’t want stored in logs.

There are real-world examples: an engineer once accidentally pasted corporate source code into a chatbot, prompting some companies to ban AI tools for work use — a reminder that leaks happen and can be costly.

Quick safety tips: treat ChatGPT like any online service — avoid sharing identifying facts, don’t paste secrets, use temporary/incognito chats when possible, and regularly clear history if you’ve shared sensitive content. For medical, legal or safety-critical issues, consult qualified professionals — AI is not a substitute.

🚀 OpenAI Launches ChatGPT Atlas (Oct 21, 2025)
OpenAI introduced ChatGPT Atlas, a browser built around an “agentic” AI. It was announced as a reimagining of the browser, but the first release feels experimental and unfinished.

⚠️ Key Problems
• 🐢 Slow & unreliable: Agent mode often freezes, drains batteries (especially on Macs), and struggles with multi-step tasks.
• ⚡ Performance gap: Competing agent browsers (like Comet from Perplexity) are up to 8× faster and more stable.
• 💰 Paywall: Advanced agent features are limited to Plus/Pro subscribers — free users only get basic tools.

🔒 Security & Privacy Concerns
• 🧠 Prompt injection risk: Hidden instructions on web pages can trick the agent into unwanted actions or leaking data. OpenAI admits this is a known issue.
• 📋 Memory tracking: The “browser memory” logs user actions for personalization; optional, but most users won’t tweak settings.

💡 Strengths & Good Use Cases
• 🧩 ChatGPT sidebar: Excellent for summarizing pages, editing text, and answering questions quickly.
• ✍️ Simple workflows: Works well for drafting, rewriting, or making basic presentations — ideal for existing ChatGPT fans.

🚫 Limitations
• 🛒 Not reliable yet for shopping, booking, or finance tasks.
• 🕓 Often fails at time-sensitive or aesthetic decisions.
• 🍏 Currently macOS-only; Windows and Android versions “coming soon.”

✅ Bottom Line
Atlas is an ambitious experiment blending web browsing with AI automation. It shines in light, structured tasks but isn’t ready to replace your main browser yet. For now, it’s a preview of what agentic browsing could become, not a finished product.

@Chatgptt

🔥 DeepSeek-OCR: A New Way to Compress Text via Visual Encoding

DeepSeek has unveiled DeepSeek-OCR, a breakthrough model that converts documents into visual representations before decoding them back into text — cutting token usage by 7–20× while keeping accuracy as high as 97% at less than 10× compression.

How it works:
• DeepEncoder compresses text, tables, and charts into compact visual embeddings.
• DeepSeek3B-MoE-A570M — a 3B-parameter mixture-of-experts decoder — reconstructs the original text from the compressed visuals.

Why it matters:
This approach handles structured data far more efficiently than token-based models, allowing longer context processing with drastically lower compute costs.

Use cases:
• Finance — parsing reports and tables.
• Science/healthcare — analyzing research papers and structured forms.
• Enterprise search — large-scale document indexing with minimal resource use.

Open source:
DeepSeek-OCR is available on GitHub and Hugging Face, ready for experimentation and integration.

🚀 In short: DeepSeek-OCR shows that “compressing meaning visually” can make language models faster, cheaper, and smarter with big document sets.

@DeepSeekk

🚀 OpenAI Wants to Turn ChatGPT into… an Operating System
ChatGPT’s product lead, Nick Turley, said the company envisions the chatbot as a platform or even an “OS” with its own app ecosystem — much like how web browsers became the main workspace for online services.

• Platform + App Store: Third-party services (like Expedia, DoorDash, Uber, etc.) can integrate directly into ChatGPT — a clear path for transactions and monetization.
• Hardware & design ambitions: OpenAI is reportedly exploring a browser and potential collaboration with Jony Ive’s ex-Apple team — hinting at a hardware-software ecosystem.
• Privacy & data: Third-party developers may need some user data access; OpenAI promises “minimal collection” and a concept of shared memory, though details remain vague.

💡 Huge potential — a seamless interface and unified platform for services and microtransactions.
But the biggest challenges lie in user trust, transparent monetization, and data security.